A practical, no-nonsense guide for founders, engineers, and AI teams deciding which open source or open-weight models are actually worth testing — by workload, benchmark profile, license fit, and hardware reality.
Start with a model that fits your available memory and test it on work you can score. The useful comparison is task completion, output quality, latency, and operating cost at a known context length. Model size and launch charts help narrow the options; they do not settle that comparison.
Qwen3.8-27B and Muse Glimmer-30B remain useful single-GPU candidates. Their named Q4 downloads are roughly 16–17 GB before vision files, caches, and runtime memory. A 24 GB card is a starting point for modest contexts, not a promise that every configuration fits.
Qwen3.8-Flash-Next adds a 125B backbone with 6B active parameters, plus 51B of n-gram embeddings and 4B of prediction weights. GLM-5.3 now has downloadable weights, replacing its previous pending status here. GLM-5.3-Flash adds a separate 320B / 18B-active multimodal architecture. All require substantially more memory than the single-card models.
Flash-Next uses Qwen Community License 1.0. Its commercial-use gate for model-as-a-service and AI-work-assistant businesses has no revenue minimum, unlike the $50M condition on the separate 2.4T checkpoint. Internal use has an explicit exception. The Apache-licensed 27B model is a different legal choice.
GLM-5.3 uses a custom license with a security-review requirement for model-as-a-service operators whose aggregate revenue, including affiliates, exceeds $10B over twelve months. GLM-5.3-Flash uses MIT. Neither a family name nor a quantizer’s metadata replaces the upstream license.
A terminal benchmark score describes a model running inside a particular harness with a particular time and token budget. This revision removes stale independent-index rankings and retains a smaller set of attributable publisher results. S5 Labs has not benchmarked these models; use the scores to form an evaluation shortlist.
Count every shard, vision projector, and optional drafting model. Then allow for caches, intermediate buffers, the operating system, and concurrent requests. The machine guide explains that budget; active MoE parameters describe per-token computation and do not tell you how much memory the checkpoint occupies.
Added Kolibri-1 with its checkpoint license, total and active parameter counts, context recommendation, and FP8 weight-memory caveat. This targeted update follows the September 5 guide review; other model recommendations have not all been reverified today. The guide summarizes published sources, not S5 Labs inference benchmarks.
This matrix is designed to help teams choose a sensible starting point instead of trying everything at once.
| Use case | Start here | Why | Hardware reality | Watch-out |
|---|---|---|---|---|
| General writing, chat, summaries, RAG | Qwen3.8-27B or Muse Glimmer-30B | Both are permissively licensed vision-language candidates with tool support. Compare their answers and tool calls on your own documents. | About 16–17 GB for the named Q4 language-model files, plus vision and runtime memory. Start with a 24 GB GPU at modest context, or a unified-memory machine with adequate free memory. | Qwen defaults to xhigh reasoning effort; try medium or low for interactive work. Muse Glimmer includes a separate usage policy. |
| Coding assistant for real development work | Qwen3.8-27B on one card; Qwen3-Coder-Next on a workstation | Qwen3.8-27B is the smaller general model; Qwen3-Coder-Next is the dedicated 80B / 3B-active coding MoE. | Coder-Next’s official Q4_K_M shards total 48.41 GB before runtime allocations. Prefer more memory or explicit offload; a 48 GB GPU is not a complete runtime budget. | Check the exact quantization, supported tool format, and repository-level tests. Coder-Next supports non-thinking mode only. |
| Tool-using agents and workflow automation | Muse Glimmer-30B or Qwen3.8-27B | Evaluate structured output, tool selection, and recovery from failed actions before expanding permissions. | For larger deployments, compare DeepSeek-V4-Flash-0731 and GLM-5.3-Flash. Qwen3.8-Flash-Next is another candidate after checking its commercial-use terms. | Treat retrieved pages and tool responses as untrusted input. A guard model does not replace scoped permissions, sandboxing, or approval boundaries. |
| All-in-one open-weight flagship (reasoning + coding + multimodal) | GLM-5.3-Flash or Mistral Medium 3.5 | GLM-5.3-Flash brings native multimodal input under MIT. Mistral Medium 3.5 is a dense alternative with custom commercial terms. | Multi-GPU or large-memory infrastructure; compare the complete checkpoint and serving recipe. | Flash is not a small local model: the publisher describes 320B total and 18B active parameters. |
| Top-end open reasoning and large-scale inference | DeepSeek-V4.1-Flash, GLM-5.3, or GLM-5.3-Flash | These are larger coding and agent candidates with published serving recipes. GLM-5.3 is now downloadable; Flash is a separate multimodal model. | Large-memory, multi-GPU deployments. Test quantized checkpoints and concurrency separately. | GLM-5.3 is custom-licensed; Flash and the named DeepSeek releases use MIT. Hosted API capabilities and prices must be checked separately. |
| Omnimodal: speech + vision + text in one model | Qwen3-Omni-30B-A3B-Instruct | Verified downloadable Apache-licensed model with text, image, audio, and video input and text or speech output. | Check the official serving recipe and enabled thinker/talker components. MiniCPM-o 4.5 is a smaller alternative. | The Instruct and Thinking variants have different output capabilities. Do not assume a newer hosted omni product has the same downloadable weights. |
| Vision-language understanding | Gemma 4 (E4B, 12B Unified, or 26B A4B) or Qwen3.8-27B | Gemma 4 offers several multimodal sizes under Apache 2.0. Qwen3.8-27B accepts image and video input; evaluate document accuracy on representative scans and layouts. | Compare the exact Gemma variant and quantization. A 24 GB GPU is a candidate for some Q4 configurations at modest context, not a family-wide guarantee. | For true video grounding, pointing, and counting, prefer Molmo 2. It is purpose-built for grounded visual reasoning. |
| Edge or low-footprint multimodal | MiniCPM-V 4.6 (1.3B) or Gemma 4 E2B | Edge-friendly visual models for phones, IoT, and lightweight servers. Apache 2.0. | Runs on consumer hardware with 4–8GB VRAM, sometimes CPU-only. | Pick for footprint rather than absolute accuracy; the capability ceiling is real. |
| Image generation | SDXL or FLUX.1-schnell | SDXL remains the mature ecosystem baseline. FLUX.1-schnell is the fast, permissive 12B alternative under Apache 2.0, and FLUX.2 [klein] 4B (Apache 2.0, January 2026) is Black Forest Labs' newer permissive small model. | SDXL is happiest around 12GB VRAM; FLUX.1-schnell needs more for full quality; klein 4B is reported to run in about 13 GB. | FLUX licensing is checkpoint-specific: schnell and klein 4B are Apache 2.0, FLUX.1-dev and klein 9B are non-commercial. Check newer hosted FLUX releases separately for weight availability. |
| Image editing, control, inpainting | Diffusers + ControlNet + SAM 3.1 + LaMa | Editing is a stack problem, not a single-model problem. SAM 3.1 is Meta's current segmentation line, superseding SAM 2. | Consumer GPUs handle most workflows. | Workflow quality depends on masks, conditioning, and operator skill. |
| Speech recognition | Whisper large-v3 or Whisper turbo | Large-v3 is the larger Whisper baseline; turbo is the faster official derivative for lower-latency transcription. Qwen3-ASR (Apache 2.0, 52 languages/dialects, forced aligner) is a newer open-source alternative worth evaluating for multilingual workloads. | Consumer GPUs handle both; large-v3 needs ~10GB VRAM, turbo ~6GB. | Accuracy varies significantly across the long tail of languages. Verify for your domain. Qwen3-ASR doesn't replace Whisper outright; test on your actual language mix. |
| Speech synthesis (TTS) | Qwen3-TTS (0.6B or 1.7B) or Chatterbox | Qwen3-TTS is Apache 2.0, multilingual, supports voice cloning, and has strong community adoption. Chatterbox (Resemble AI) offers Turbo (350M) for low-latency English and Multilingual V3 for broader language coverage. For high multilingual quality, Voxtral TTS (4B) is a strong open-weight option but carries CC-BY-NC weights. | Qwen3-TTS and Chatterbox run comfortably on consumer GPUs; viable for real-time agent voices. | Voice cloning has ethical and legal implications. Confirm consent and disclosure rules. Voxtral TTS is non-commercial under its CC-BY-NC weight terms; confirm before any commercial use. |
| Safety classification and guardrails | Shieldstral or gpt-oss-safeguard | Shieldstral (Mistral, 3B, Apache 2.0) classifies text and images against a policy you write in plain language at inference time. gpt-oss-safeguard (OpenAI, 120b and 20b, Apache 2.0) reasons over a developer-supplied policy and explains its verdicts. | Shieldstral runs on a single 16 GB NVIDIA GPU in BF16; gpt-oss-safeguard-20b has the same 21B footprint as gpt-oss-20b (16 GB), the 120b needs one 80 GB GPU. | A guard model is a component, not a policy. Measure false-positive rates on your own traffic before you let it block anything, and keep it in front of the agent, not only the chat box. |
| Open video experiments | Wan2.2 or Open-Sora 2.0 | Wan2.2 (Apache 2.0) has a 5B 720p TI2V path that is practical on strong consumer GPUs and remains the open-weight video model; the Wan 2.5–3.0 line is API-only. Open-Sora 2.0 (11B, Apache 2.0) is the openly trained R&D counterpart. LTX-2.5 (22B, joint audio and video, free for commercial use under $10M in revenue) is the new option if you need synchronized sound. | Wan2.2 TI2V-5B needs at least 24 GB VRAM with model offload (an RTX 4090, per the README); the 27B A14B models want 80 GB or multi-GPU. | Video quality, latency, and consistency remain uneven. Treat as R&D, not production-default. MiniMax H3's license excludes US, EU, UK and Korean users, so it is not an option here. |
Choose the smallest model that reliably completes your real task with the right output shape. Then move up only if the gains are measurable.
This page separates fully open releases from open-weight releases, license-restricted releases, and models that have only been announced — because the market still collapses all four into the same marketing label.
The OSI Open Source AI Definition calls for the data information, code, and parameters needed to study and modify a system. It does not simply require an Apache label or necessarily publication of every raw training record. OLMo is useful here because Ai2 publishes substantial training artifacts alongside its weights.
You can download and run the weights, but the full training data and recipe are not completely reproducible. This is where most high-performing "open" models sit today, including families under permissive licenses (Apache 2.0 / MIT / OpenMDW) and those under community, modified, or revenue-triggered licenses.
Custom terms vary considerably. Llama 4 has a community license and separate acceptable-use policy. Qwen3.8-Flash-Next requires a separate commercial license for certain AI businesses regardless of revenue; GLM-5.3 has a much higher, narrowly defined security-review threshold. Non-commercial weight licenses such as Voxtral TTS are a separate constraint.
A hosted model announcement does not establish that a usable checkpoint exists. Check for actual weight files, a model card, and a license before adding a model to a self-hosting plan. GLM-5.3 has now cleared that artifact check and moved into the comparison table.
Candidates for text, code, vision, speech, and video workloads. Each row identifies the exact checkpoint’s terms and deployment requirements. Status badges help distinguish permissive weights from custom or restricted releases.
Weights under Apache 2.0, MIT, or another permissive license. Check attribution, usage policies, and third-party components before deployment. A permissive weight license alone does not establish compliance with the OSI Open Source AI Definition.
Weights, code, and substantial training information are available for research and inspection. Reproducibility still depends on compute, tooling, and the completeness of those artifacts.
Downloadable weights with additional conditions, such as business-use restrictions, revenue thresholds, or acceptable-use policies. Read the exact checkpoint license.
Downloadable weights with non-commercial or other significant use restrictions. Confirm that the intended use is permitted.
A model with a press release and no downloadable weights or license. It does not belong in an evaluation shortlist until the files exist.
| Model | Status | License | Parameters | Key capabilities | Inference target | Maturity |
|---|---|---|---|---|---|---|
| Kolibri-1 | Permissive weights | Apache 2.0 | 78B total / 3.46B active per token | German/English reasoning and tools; 1,048,576 maximum context; publisher recommends ≤262,144 | About 78 GB FP8 weights alone; add caches and runtime memory | Released October 3, 2026; source reviewed October 5, no S5 Labs runtime test |
| Qwen3.8-27B | Permissive weights | Apache 2.0 | 27B language model plus vision components | General text, coding, agents; image and video input; 262K context | Named community UD-Q4_K_M: 16.46 GB + 0.93 GB vision; add runtime headroom | Candidate for local evaluation; community quantizations |
| Muse Glimmer-30B | Permissive weights | Apache 2.0 + separate Meta usage policy | ~29.6B dense + 1.8B vision encoder | Tool-calling agents, search, image input; 131K context | Official Q4: 16.76 GB + 1.40 GB vision; optional drafter adds 1.63 GB | First-party quantized artifacts and local runtime guidance |
| Qwen3.6-35B-A3B | Permissive weights | Apache 2.0 | 35B total / 3B active | General text, coding, agentic workflows, 262K context | Prosumer local — 32–64GB unified or 24GB+ GPU with sufficient runtime headroom | Production-leaning; the MoE fallback |
| Qwen3-Coder-Next | Permissive weights | Apache 2.0 | 80B total / 3B active | Coding agents, tool use, long-horizon reasoning, 262K context | Official Q4_K_M: 48.41 GB before cache and runtime; larger memory or offload | Dedicated coding candidate; non-thinking mode |
| Gemma 4 31B / 12B Unified | Permissive weights | Apache 2.0 | 30.7B dense or 11.95B unified | Text, coding, image; audio on 12B Unified | 12B on a 16 GB laptop (per Google); 31B should fit a 24–48 GB card at Q4 | Production-leaning; official QAT GGUFs |
| gpt-oss-120b / 20b | Permissive weights | Apache 2.0 + usage policy | 117B / 5.1B active · 21B / 3.6B active | Text reasoning, tool use, structured output | Single 80GB GPU · 16 GB device | Production-leaning |
| Nemotron 3.5 Lightning | Permissive weights | OpenMDW-1.1 | 30B total / 3B active | Fast local agents and sub-agents; 1M context nominal | Check the selected NVFP4/BF16 artifact and NVIDIA serving recipe | Hybrid architecture; evaluate speed and quality on target hardware |
| Qwen3.8-Flash-Next | Open weight, custom terms | Qwen Community License 1.0; business-use gate has no revenue floor | 125B backbone / 6B active + 51B n-gram + 4B MTP | Text, image, video; 262K native context | Roughly 180B checkpoint; high-memory deployment | Downloadable; verify license before commercial evaluation |
| GLM-5.3 | Open weight, custom terms | Custom GLM-5.3 License; conditional security review | Approximately 753B checkpoint | Text coding and long-horizon agents | Large-memory / multi-GPU; official FP8 and BF16 repositories | Weights available; previously listed here as pending |
| GLM-5.3-Flash | Permissive weights | MIT | 320B total / 18B active (publisher) | Native multimodal input; hybrid sparse and linear attention | Large-memory / multi-GPU; recipe-specific quantization | Separate architecture from GLM-5.3; released weights |
| DeepSeek-V4.1-Flash | Permissive weights | MIT | 552B backbone + 196B Engram; 8B active prefill / 16B decode | Native image and text input, text output; 1M context | Large-memory / multi-GPU; count all checkpoint components and runtime overhead | Released September 10, 2026; separate architecture from V4 Flash |
| DeepSeek-V4-Flash-0731 | Permissive weights | MIT | 284B total / 13B active | Agentic coding and tool use, 1M context, DSpark speculative decoding | Multi-GPU (~167 GB FP4/FP8); reference 4×GB300 node | Official release (Jul. 31, 2026), supersedes the April preview |
| DeepSeek-V4-Pro-0813 | Permissive weights | MIT | 1.6T total / 49B active | Frontier reasoning and agentic behavior, 1M context | Datacenter — 4×GB300 reference deployment | GA (Aug. 13, 2026), supersedes the April preview |
| Inkling | Permissive weights | Apache 2.0 (ungated download) | 975B total / 41B active | Natively multimodal (text, image, audio in) frontier-scale reasoning | Datacenter — 1.91 TB of weights; Inkling-Small (276B / 12B) at ~532 GB | New (July 15, 2026) — evaluate before trusting |
| Molmo2-8B | Permissive weights | Apache 2.0 (data-use caveats) | 8B | Image, video, grounding, pointing, tracking | Workstation / server | Research + production pilots (Dec. 2025; full training code Mar. 2026) |
| MiniCPM-o 4.5 | Permissive weights | Apache 2.0 | 9B | Real-time omnimodal — text, vision, speech | 19 GB BF16 / 11 GB INT4 (per OpenBMB); quantized local | Production-leaning for edge/omni |
| Whisper large-v3-turbo | Permissive weights | MIT (code and weights) | 0.8B | ASR, low-latency transcription | CPU, modest GPU (~6GB VRAM), or edge device | Mature / production standard |
| Qwen3-TTS 1.7B / 0.6B | Permissive weights | Apache 2.0 | 1.7B or 0.6B | Multilingual TTS, voice cloning, streaming | Consumer / prosumer GPU | Production-leaning |
| Wan2.2 TI2V-5B | Permissive weights | Apache 2.0 | 5B (family also has 27B / 14B-active MoE models) | Text- and image-to-video, 720p at 24 fps | 24 GB GPU with model offload (RTX 4090 per the README) | Advanced prototyping / R&D; open-weight Wan model |
| OLMo 3.1 32B-Think | Fully open | Apache 2.0 + published data and recipe | 32B dense | Auditable reasoning with full model-flow traceability | Prosumer local — 32–64GB | Mature for research and governed evaluation |
| OLMo Hybrid 7B | Fully open | Apache 2.0 + published data and recipe | 7B | Hybrid Gated DeltaNet architecture research; 65K context | Consumer GPU | Research (Mar. 5, 2026); no RL-final chat checkpoint yet |
| Llama 4 Scout | Open weight, custom terms | Llama 4 Community License + AUP | 109B total / 17B active | Multimodal text+image, 10M context | INT4 on a single H100 or a 128GB unified box | Production-leaning; Meta's newer open weights are Muse Glimmer |
| Mistral Medium 3.5 | Open weight, custom terms | Modified MIT — no rights above $20M monthly revenue | 128B dense | Reasoning, coding, multimodal, agents; 256K context | ~4-GPU self-hosting | Production-leaning |
| Kimi K3 | Open weight, custom terms | Custom Kimi K3 License; conditional MaaS agreement and UI attribution | 2.8T total / 104B active | Native vision, 1M context, MXFP4 weights | Datacenter — 1.56TB across 96 shards; 64+ accelerators per Moonshot | Weights July 27, 2026 — verify license fit first |
| Qwen3.8-2.4T-A95B | Open weight, custom terms | Qwen3.8-Max License — $50M trigger for MaaS / AI-work-assistant businesses + UI attribution | 2.4T total / 95B active | Text-only, thinking-only checkpoint of the hosted Qwen3.8-Max; 262K native context | Datacenter — 2.5 TB FP8 / 4.89 TB BF16; GB300 NVL72 reference | Weights Aug. 12, 2026 — the checkpoint is not the hosted product |
| Voxtral TTS | Restricted / non-commercial | CC-BY-NC 4.0 (inherited from reference voices) | 4B | High-quality multilingual speech synthesis | Consumer / prosumer GPU | Non-commercial use only |
| Muse Spark 1.2 (weights promised) | Announced, not released | None published | Not disclosed | Hosted model; not included as a downloadable recommendation | No checkpoint verified in this review | Re-check publisher artifacts before evaluation |
Check the upstream license and any usage policy for the exact files you plan to deploy. Qwen3.8-27B, Flash-Next, and the 2.4T checkpoint have different terms; GLM-5.3 and GLM-5.3-Flash differ too. Download size is separate from runtime memory. Pair the inference-target column with our local LLM machines guide before choosing hardware.
Open AI is no longer one category. The ecosystem now includes general-purpose LLMs, code-specialized models, multimodal models, image generators, image-editing stacks, and increasingly capable video systems.
| Category | Best for | Start here | Move up to | Sweet spot | Watch-outs |
|---|---|---|---|---|---|
| Writing / general LLM | Chat, drafting, summarization, RAG, internal copilots | Qwen3.8-27B or Muse Glimmer-30B | DeepSeek-V4-Flash-0731, GLM-5.3-Flash, Mistral Medium 3.5 | Qwen3.8-27B and gpt-oss-20b offer smaller starting points; Qwen3.6-35B-A3B remains the MoE fallback when decode speed on a laptop matters more than peak scores. | License terms (Llama 4 EU multimodal limits, revenue triggers on Kimi K3, Qwen3.8-2.4T-A95B and Mistral Medium 3.5) and quantization quality matter more than leaderboard hype. |
| Coding | PR assistance, code generation, refactors, test writing, coding agents | Qwen3.8-27B or Qwen3-Coder-Next when sufficient memory is available | GLM-5.3, DeepSeek-V4-Pro-0813, or Qwen3.8-Flash-Next after a license review | Qwen3.8-27B is a candidate for a single card; Qwen3-Coder-Next (80B / 3B active, 262K context) when the workstation has the memory. Leanstral 1.5 (119B / 6B active, Apache 2.0) is the specialist for Lean 4 formal verification. | Do not deploy without tests, sandboxing, and dependency/security review. Self-graded agent scores need an outside reproduction before they go into a procurement deck. |
| Agents | Tool use, workflow automation, multi-step task execution | Muse Glimmer-30B or Qwen3.8-27B | GLM-5.3-Flash or DeepSeek-V4-Flash-0731; compare with Qwen3.8-Flash-Next if its terms fit | Measure end-to-end completion, retries, and time spent waiting on tools. Faster token generation alone does not establish a faster agent. | JSON breakage, tool misuse, cascading failures, and prompt injection are the real bottlenecks. Confirm tool-calling support on the exact checkpoint before committing, and put a guard model in front of anything that reads untrusted input. |
| Multimodal / VLM | Document understanding, image Q&A, visual agents, OCR-heavy workflows | Gemma 4 (E4B, 12B Unified or 26B A4B) or Qwen3.8-27B | GLM-5.3-Flash or Qwen3.8-Flash-Next, with a separate license and memory review | Gemma 4 26B A4B and Qwen3.8-27B are the practical local starting points; Qwen3.8-27B adds video input and the stronger document-parsing score. | Grounding mistakes and OCR hallucinations still require checks; for true visual grounding, prefer Molmo 2. |
| Omnimodal (text + vision + speech) | Voice-first assistants, real-time audio-visual reasoning, streaming speech generation | MiniCPM-o 4.5 (9B) for local real-time | Qwen3-Omni-30B-A3B-Instruct for a downloadable thinker/talker model; Gemma 4 12B when audio input with text output is enough | MiniCPM-o 4.5 gives most teams a real-time omni model that fits on prosumer hardware (19 GB BF16, 11 GB INT4 per OpenBMB). | Some serving stacks still need patched support; streaming speech latency is the binding constraint. |
| Video grounding / pointing | Video understanding, object pointing, tracking, multi-image reasoning | Molmo 2 (4B, 8B, O-7B variants) | Custom Molmo 2 fine-tunes on domain data | One of the few open families with video grounding and pointing under Apache 2.0, and since March 2026 the full training code and stage checkpoints; note the third-party dataset restrictions in the model card. | Not a general chat model. Pair it with a chat-capable LLM for conversational interfaces. |
| Edge multimodal | Phones, IoT, lightweight servers, on-device VLM workloads | MiniCPM-V 4.6 (1.3B) or Gemma 4 E2B | MiniCPM-o 4.5 or Gemma 4 E4B when more capability is required | MiniCPM-V 4.6 is a small visual model to evaluate on constrained devices. | Capability ceiling is real. Do not expect 27B-class reasoning at 1.3B. |
| Embeddings / retrieval | RAG indexes, semantic search, retrieval over mixed text, image, video and audio, on-device search | EmbeddingGemma 2 (one checkpoint, 270M text-only to 740M full) | Qwen3-VL-Embedding-2B (Apache 2.0; text, image and video, no audio) when image and video quality matters more than size. At about 2.1B parameters it self-reports 73.2 on MMEB v2 (73.25 in the leaderboard score file) against EmbeddingGemma 2's 59.01, which Google submitted for the text-plus-vision load. | Load only the encoders you need: 270M for text only, 440M for text plus images (Google's developer guide also lists video), 570M for text plus audio. Google says that in sentence-transformers all four loads share one 768-dimensional space, so a text-only query model can search vectors from the full model. S5 Labs has not tested that, and the card says encoder-omission loading differs among model libraries. | Google tags EmbeddingGemma 2 Apache 2.0 in repo metadata and states it on the card and in its launch posts, but the repo has no LICENSE file and the card still cites the Gemma Prohibited Use Policy. EmbeddingGemma 2 scores are Google's own, and MTEB had not accepted the submission at launch. Changing embedding models means re-embedding the index. Google says 128 dimensions suits text-only work best; its table shows MMEB at 77.4% of the 768-d score. Retrieve per modality and merge until you have checked cross-modal score calibration. |
| Image generation | Concept art, marketing assets, ideation, product visuals | SDXL | FLUX.1-schnell or FLUX.2 [klein] 4B (both Apache 2.0), or other FLUX variants when licensing permits | SDXL has an established ecosystem; the permissive FLUX checkpoints when you need speed and a clean license. | Typography and exact prompt fidelity still need workflow iteration. FLUX licensing is checkpoint-specific. |
| Image editing | Inpainting, control, masking, pose/depth guidance, product edits | ControlNet + SAM 3.1 + LaMa + Diffusers | Project-specific editing stacks with custom masks and pipelines | Editing quality comes from how the stack is designed rather than from any single checkpoint. | Commercial rights differ across base checkpoints and extensions. |
| Speech recognition | Transcription, translation, voice interfaces, audio understanding | Whisper large-v3 | Whisper turbo for faster transcription; Qwen3-ASR for another multilingual option; Qwen3-Omni for combined audio and text reasoning | Whisper large-v3 supports transcription and translation; turbo is the official low-latency derivative. | Accuracy varies significantly across the long tail of languages. Verify for your domain. |
| Speech synthesis (TTS) | Voice agents, narration, dubbing, expressive synthesis | Qwen3-TTS (Apache 2.0, 0.6B or 1.7B, multilingual, voice cloning) or Chatterbox-Turbo (350M) / Chatterbox-Multilingual V3 (500M) | Voxtral TTS (4B, CC-BY-NC, non-commercial only) for top multilingual quality and voice adaptation | Qwen3-TTS is the permissive default; step up to Voxtral TTS only when multilingual quality is the priority and non-commercial terms are acceptable. | Voice cloning has ethical and legal implications. Confirm consent rules before deploying. Voxtral TTS is CC-BY-NC 4.0: non-commercial use only. |
| Safety / guardrails | Input and output moderation, policy enforcement in front of agents, refusal detection | Shieldstral (3B, Apache 2.0, text + image, 16 GB GPU) | gpt-oss-safeguard 20b or 120b (Apache 2.0) when you want the classifier to reason over a long written policy and explain itself | Shieldstral is small enough to colocate with the application; its policy is a prompt, so changing it is a deploy, not a retrain. | Vendor moderation scores are within a point of each other on text (Shieldstral 88.1 vs Qwen3Guard-8B 88.2 on WildGuardTest, per Mistral). The differentiator is your false-positive rate on your traffic, which no card publishes. |
| Video generation | Short exploratory clips, motion concepts, early creative prototyping | Wan2.2 TI2V-5B (Apache 2.0) | Larger Wan2.2 A14B variants, Open-Sora 2.0, or LTX-2.5 for synchronized audio and video after reviewing its Community License | Today, open video is a prototyping tool more than a production default. | Temporal flicker, identity drift, and long render times remain common. Do not assume newer hosted Wan versions have downloadable checkpoints; do not assume a repo exists because a version number does. |
| Video editing | Interpolation, inpainting, retiming, experimental edit pipelines | RIFE, ProPainter, Wan2.2 VACE-style workflows | Custom pipelines for domain-specific video tasks | Use specialized tools rather than expecting one general model to handle everything. | Workflow complexity is high; results are sensitive to clip quality and masking. Some tools (e.g., ProPainter) ship under research-only S-Lab terms. Verify license fit. |
The biggest change from 2024 to 2026 is not just raw model quality. It is the breadth of credible open options across text, coding, multimodal, and media generation.
Short, opinionated notes on the families that matter most right now — what each is suited to, and where it stops being the right tool.
Aleph Alpha’s model card describes a 78B MoE with 3.46B active parameters per token. Its checkpoint license is Apache 2.0. The card lists a 1,048,576-token maximum but recommends no more than 262,144 for serving efficiency and complex tasks. The roughly 78 GB FP8 weight footprint excludes runtime allocations; active parameters do not make this a 3.46B-sized download. Evaluate German and English tasks separately on your intended serving setup.
Qwen’s card describes a 27B language model with image and video input and 262,144 native context. The Unsloth UD-Q4_K_M language-model download is 16.46 GB, plus a separate vision projector. That makes modest-context testing on a 24 GB card plausible, but context and runtime settings still determine fit. Qwen exposes low, medium, and xhigh reasoning effort; xhigh is the default. Start with a budget that matches the task. Our launch analysis provides the earlier release context.
Meta publishes a 16.76 GB Q4 language-model file, a 1.40 GB vision file, and an optional 1.63 GB DFlash drafter. Count the components you enable rather than treating “17 GB” as total runtime memory. The model accepts text and images, with a 131,072-token context specified in the card. Apache 2.0 is accompanied by a usage policy. Our Glimmer article covers the launch.
The card separates a 125B backbone with 6B active parameters from 51B of n-gram embeddings and 4B of multi-token-prediction weights. The full checkpoint is roughly 180B parameters. It supports image and video input with 262,144 native context. Its new Community License 1.0 requires a separate license for commercial use by specified AI businesses regardless of revenue, with an internal-use exception. That makes it a different procurement decision from Qwen3.8-27B.
GLM-5.3 has moved out of this guide’s pending category. The published FP8 repository is roughly 753B parameters. GLM-5.3-Flash instead uses a new 320B / 18B-active multimodal architecture and MIT terms. Both cards document low, high, and max reasoning effort; max is the default. Use the matching serving recipe and template. The Flash name describes a different efficiency tradeoff, not a desktop-sized model.
The MIT-licensed checkpoint accepts images and text and generates text with up to 1M context. DeepSeek separates its 552B backbone from 196B of Engram conditional memory, with 8B active parameters during prefill and 16B during decode. The backbone count is not the complete checkpoint or runtime memory requirement. Evaluate the matching serving implementation and all required components. Our release analysis covers the API migration; changes to hosted aliases do not remove the older downloadable V4 checkpoints below.
The MIT-licensed V4-Flash-0731 and V4-Pro-0813 releases remain larger-model candidates. Their cards describe 284B / 13B-active and 1.6T / 49B-active architectures respectively, with 1M context. Use the complete suffixed repository name to distinguish them from preview artifacts. Published agent scores are tied to DeepSeek Harness. Compare the runtime recipe and actual checkpoint size before deciding whether to self-host; hosted API prices and features are a separate decision.
The downloadable checkpoint is text-only and thinking-only, while the hosted Qwen3.8-Max product has additional features. Its card labels benchmark results with the hosted product’s name, so do not assume feature parity. The custom license has an aggregate-revenue condition for specified AI businesses, separate UI-attribution conditions, and an internal-use exception. It differs from both Apache-licensed Qwen3.8-27B and Flash-Next’s Community License 1.0.
Mistral describes a 128B dense model with 256K context and multimodal input. Its card reports SWE-bench Verified 77.6%; that publisher result needs its harness context. The modified MIT license includes a $20M global consolidated monthly-revenue condition. Check the actual license and serving requirements before treating this as an interchangeable alternative to permissively licensed models.
Qwen’s dedicated coding MoE has 80B total and 3B active parameters, 262,144 native context, and non-thinking output. The official Q4_K_M shards total 48.41 GB. Sparse activation reduces computation per token, but the other experts still occupy storage and generally memory. Add cache and runtime headroom or plan explicit offload; the file size alone does not establish that a 48 GB GPU fits.
This Apache-licensed 35B / 3B-active model remains useful for comparing a small MoE against a dense local model. Its native context is 262,144 tokens. Measure prefill and decode separately on your machine: active-parameter count alone does not establish a speed multiplier.
Thinking Machines publishes Inkling with Apache 2.0 metadata and text, image, and audio input. Its 975B / 41B-active architecture is a large deployment commitment. Evaluate it for a specific workload rather than inferring reliability from openness or size. The smaller Inkling-Small release is another candidate when the full checkpoint is impractical.
Moonshot’s card describes a 2.8T / 104B-active model with MXFP4 weights. The custom license’s $20M test uses aggregate revenue of the licensee and affiliates operating a model-as-a-service business, not only revenue earned from that model. Internal-use and official-provider exceptions are explicit, and UI attribution has separate thresholds. Read those definitions before choosing a hosting arrangement.
Scout is a 109B / 17B-active multimodal model with a published 10M-token context ceiling. Meta’s quantized single-H100 guidance does not mean the maximum context fits in that configuration. The community license and separate acceptable-use policy include conditions that differ from Muse Glimmer’s terms.
Google’s current card lists E2B, E4B, 12B, 26B A4B, and 31B variants under Apache 2.0. Text and image input span the family; audio input is supported on E2B, E4B, and 12B. Context limits and architecture differ by size. The smaller models use 128K context and the medium models support 256K. Pick a checkpoint for its modalities and memory needs rather than treating the whole family as one model.
OpenAI’s cards describe text-only models with tool and structured-output support. The published deployment targets are a single 80 GB GPU for 120b and a 16 GB device for 20b. Those are vendor targets, not guarantees for every context or backend. The safeguard variants classify text against a supplied policy. Read the usage policy alongside Apache 2.0 and test the required output format.
Ai2 publishes model, training, and data artifacts for OLMo. OLMo 3.1 32B-Think provides an auditable reasoning checkpoint; OLMo Hybrid explores a mix of Gated DeltaNet and attention layers. These releases are useful when inspecting training matters as much as answer quality. Reproducing training still requires the relevant data, software, and compute.
Ai2’s Molmo 2 family supports image and video understanding, including pointing and tracking. The model card distinguishes its Apache 2.0 artifacts from conditions on some third-party training datasets. Review those terms for the activity you intend, especially training or redistributing data. Grounding quality should be checked on your images and videos.
OpenBMB combines text, vision, and speech capabilities in a 9B model. Its published memory table lists 19 GB in BF16 and 11 GB in INT4. Those figures are a starting point for evaluating the supported runtime, streaming mode, and latency on your hardware.
OpenAI’s repository lists about 10 GB VRAM for large-v3 and 6 GB for turbo. Turbo is an optimized 809M-parameter derivative; it is not trained for translation. For speech translated into English, use a multilingual non-turbo model. Measure transcription accuracy on your language mix, noise conditions, and vocabulary. Real-time chunking is an application concern; the model does not guarantee streaming behavior.
Qwen provides Apache-licensed 0.6B and 1.7B variants with multilingual speech and voice-cloning capabilities. Check which checkpoint supports the required voice workflow. Obtain permission for voices used in cloning and evaluate pronunciation and latency on representative text.
The Voxtral-4B-TTS-2603 card identifies CC-BY-NC 4.0 terms. Treat that as a constraint on the intended use rather than assuming every internal business application is non-commercial. A permissive alternative may be simpler when commercial deployment is required.
The Wan2.2 repository documents a 5B text/image-to-video model generating 720p at 24 fps, with a 24 GB RTX 4090 example using offload. That is a specific recipe, not a general memory guarantee. Larger A14B variants require more resources. Check motion consistency, generation time, and the exact model license before incorporating a video workflow into production.
Some of the most useful tools in open AI are not foundation models. They are guard models, embedding models, conditioning architectures, segmentation models, agent harnesses and workflow utilities that sit in front of or on top of generators. They do not belong in a head-to-head leaderboard with Qwen, Gemma, or DeepSeek, but they are essential to most real production pipelines.
Mistral's 3B multimodal guard model (August 4, 2026): it scores text, images, and text-plus-image inputs against a moderation policy you write in natural language at inference time, and returns a calibrated probability in one forward pass. Fits a single 16 GB NVIDIA GPU in BF16, so it can sit next to the application rather than in a separate cluster. Mistral's own table has it within a point of Qwen3Guard-8B on text safety and well ahead of the larger OmniGuard-7B on multimodal moderation (VLGuard 97.7 vs 88.5). It is a component that sits in front of a model, not a base model in its own right.
License: Apache 2.0
OpenAI's October 2025 fine-tunes of gpt-oss that read a developer-supplied policy at inference and classify text against it with a reasoned explanation, at low, medium or high reasoning effort. Same footprints as the base models: the 120b fits one 80 GB GPU, the 20b runs in 16 GB. Built with ROOST, Discord and SafetyKit; OpenAI labels it a research preview intended for safety use cases. Heavier than Shieldstral, and the better fit when the policy is long and the verdict needs a rationale.
License: Apache 2.0 + OpenAI usage policy
Control architecture for Stable Diffusion-class workflows. Adds pose, depth, edge, and segmentation conditioning to diffusion image generation. A workflow component, not a standalone foundation model.
License: Apache 2.0 code; OpenRAIL weight distribution
Meta’s promptable segmentation model for masks and visual pipelines. The current SAM repository uses a custom license with use restrictions; do not copy the Apache license assumption from SAM 2.
License: Custom SAM License
Inpainting model for object removal and background reconstruction inside larger editing stacks.
License: Apache 2.0
Frame interpolation for slow motion and frame-rate conversion. Check the model version and test motion artifacts on the intended footage.
License: MIT (code) — verify weight distribution
Mask-aware video inpainting and object removal. Specialized for video editing; not a general-purpose video tool. S-Lab research terms rule out most commercial production uses. Confirm the license before building on it.
License: S-Lab License (research / non-commercial — verify before production)
A video-generation implementation designed to reduce inference memory use. Its code license does not relicense the Hunyuan-derived model weights used by a workflow.
License: Apache 2.0 code; inspect base-model weight terms separately
The agent runtime DeepSeek open-sourced on August 13, 2026, and the software that produced the Terminal-Bench and DeepSWE numbers on its V4 model cards ("minimal mode": one system prompt, a bash tool and a file editor). Model-agnostic, plugin-based, and labelled a developer preview with breaking changes expected. Useful in two ways: as a harness to run agents locally, and as the thing to point at when someone asks how a self-reported agent score was produced.
License: MIT
Google's October 6, 2026 embedding model. It maps text, code, images, video and audio into one 768-dimensional space. One ungated checkpoint (1.49 GB in bf16) loads as four models: 270M text-only, 440M with the image encoder, 570M with audio, or the full 740M. Matryoshka truncation to 512, 256 or 128 dimensions is supported, though Google says 128 suits text-only work best. The 8,192-token context is shared across modalities. LiteRT-LM bundles for on-device use are 165, 388 and 485 MB; run it in bf16 or fp32, because fp16 returns NaN or degraded embeddings. Scores are Google's own, and the MTEB submission had not been accepted at launch. Switching embedding models means re-embedding the corpus, and Google does not say the EmbeddingGemma 1 and 2 vector spaces are compatible.
License: Tagged Apache 2.0 in repo metadata and stated on the card (no LICENSE file; card still cites the Gemma Prohibited Use Policy)
Use these components alongside foundation models, not instead of them. Shieldstral and gpt-oss-safeguard sit in front of a model as policy filters; EmbeddingGemma 2 supplies the vectors a retrieval index searches; ControlNet adds conditioning to image generation; SAM 3.1 adds grounding to vision pipelines; FramePack, RIFE, and ProPainter extend video workflows; DeepSeek Harness runs the agent loop and, not incidentally, is how DeepSeek graded its own models.
These numbers are useful as a map, not as a verdict. Benchmark settings vary. Prompt formatting moves scores. Preference benchmarks can overstate real operational reliability. Use this as the first filter, then test on your own workload.
| Model | General | Reasoning | Coding | Notes |
|---|---|---|---|---|
| Qwen3.8-27B | GPQA Diamond 89.2 | HLE 30.8 | SWE-bench Pro 61.7; Terminal Bench 2.1 73.0 | Qwen model card. The SWE-bench Pro evaluation uses a corrected task set and Claude Code harness; it is not directly comparable with every public leaderboard. |
| Muse Glimmer-30B | MMMU Pro 74.0 | MCP Atlas 75.5; DeepSearch QA 74.6 | SWE-bench Verified 76.0; Terminal Bench 2.1 51.7 | Meta model card. Compare with the peers and evaluation settings in that card; these are not S5 Labs measurements. |
| GLM-5.3 | HLE with tools 62.5 | Toolathlon Verified 73.0 | Terminal Bench 2.1 88.2; Terminal Bench 3.0 28.3; DeepSWE v1.1 66.9 | Z.ai model card. Terminal Bench 3.0 uses three rollouts, max reasoning effort, 400K context, and a ten-hour timeout. HLE with tools is not the same test as HLE without tools. |
| DeepSeek-V4.1-Flash | Native image and text input; see publisher evaluation tables | Model-card results depend on evaluation settings | See DeepSeek Harness results and reproduction recipe | Compare matching benchmark versions and harnesses. The new architecture does not establish a universal improvement over V4 Pro. |
| DeepSeek-V4-Pro-0813 | HLE 42.7; with tools 60.0 | Toolathlon Verified 74.1 | Terminal Bench 2.1 87.9; DeepSWE 62.7 | DeepSeek model card and DeepSeek Harness. The named checkpoint is distinct from the older un-suffixed preview. |
| DeepSeek-V4-Flash-0731 | See the card for the full evaluation set | Toolathlon Verified 70.3 | Terminal Bench 2.1 82.7; DeepSWE 54.4 | DeepSeek publisher results. Do not infer equivalence with another lab’s harness from a similar score. |
| Kimi K3 | GPQA Diamond 93.5 | BrowseComp 91.2 | Terminal Bench 2.1 88.3; DeepSWE 67.5 | Moonshot model card. Coding results use its Kimi Code harness. Custom license and datacenter memory needs remain separate selection criteria. |
| Qwen3.8-Flash-Next / GLM-5.3-Flash | Published tables linked in Sources | Evaluate reasoning budget and context settings | New larger-model candidates; no combined rank assigned here | The cards expose different benchmark sets and evaluation settings. Consult each complete table rather than extrapolating a rank from model size or the Flash name. |
Use one academic snapshot table, one real-work evaluation table, and one reliability table. If a model only looks good in one of those three, it is not production-ready for your team.
Model publishers disclose unevenly, and the gaps are not random — labs publish the benchmarks they win. Every figure above should be readable as a claim with an owner, so here is who owns which kind.
The snapshot below identifies the publisher for each result. Even a familiar benchmark name can hide different task revisions, harnesses, timeouts, context lengths, and reasoning budgets. Cross-lab rows are not a controlled leaderboard.
Independent evaluators can make comparisons more consistent, but their index versions and model settings change. This guide links to external evaluations without carrying forward undated rankings or treating a composite score as a deployment recommendation.
This is a source review, not a hands-on benchmark. Hardware figures are published artifact sizes or vendor deployment guidance. We have not measured inference speed, accuracy, or reliability across this shortlist.
Z.ai’s GLM-5.3 card reports Terminal Bench 3.0 with three rollouts, max reasoning effort, a 400K context, and a ten-hour timeout. Qwen’s 27B card uses a corrected SWE-bench Pro task set. Preserve those details when comparing scores: matching benchmark names alone do not establish equivalent evaluation conditions.
The real tradeoff is not "open is better" or "closed is better." It is whether you want control, customization, and privacy enough to take on the systems burden yourself.
| Dimension | Open / open-weight | Closed ecosystem |
|---|---|---|
| Control | Self-host and customize within the checkpoint license terms. | Fastest path to strong capability with less systems work. |
| Cost model | Infrastructure, ops, and engineering replace per-token API pricing. | Usage-based pricing is simple but can become expensive at scale. |
| Privacy and data boundary | Best option when prompts, outputs, and logs must stay inside your environment. | Provider policy and retention controls matter more. |
| Customization | Adapters, quantization, routing, and domain tuning are the major advantages. | Prompting is easy; deep model customization is limited. |
| Operational burden | You own serving, evals, security, and reliability. | The provider manages infrastructure; review service terms and availability commitments. |
| Best fit | Teams with repeatable workloads, privacy needs, or platform ambitions. | Teams optimizing for speed, simplicity, and managed frontier access. |
Count weights, vision components, cache, runtime buffers, and operating-system headroom before selecting a model. If you are buying a dedicated box to run these models locally, our best local LLM machines guide compares Mac Studio, DGX Spark, Framework, HP Z2 and more by memory, bandwidth, and software stack.
What fits: Gemma 4 E2B/E4B and 12B Unified (16 GB laptop per Google), MiniCPM-V 4.6, gpt-oss-20b, Shieldstral (16 GB), EmbeddingGemma 2 (1.49 GB bf16 weights only; runtime memory not tested by S5 Labs), FLUX.2 [klein] 4B, SDXL, Whisper turbo (~6GB VRAM), Qwen3-TTS 0.6B/1.7B
Best for: Local testing, lightweight RAG, first agents, edge multimodal, image generation, on-device ASR and TTS, a guard model beside the app
Watch-outs: Named quantizations of Qwen3.8-27B and Muse Glimmer-30B are larger than this VRAM band before runtime allocations. Model downloads and system RAM do not automatically increase dedicated GPU memory.
What fits: Qwen3.8-27B (16.46 GB UD-Q4_K_M language model plus 0.93 GB BF16 vision projector), Muse Glimmer-30B (16.76 GB Q4 language model plus 1.40 GB vision file), Gemma 4, and smaller speech or vision workloads. These are candidate configurations, not tested runtime guarantees. Qwen3-Coder-Next’s Q4_K_M download totals 48.41 GB and needs additional memory or offload.
Best for: Serious private assistants, agentic coding, local omni experiments, MoE serving, open-weight video generation prototyping
Watch-outs: Budget caches, runtime buffers, and the optional drafting model. Larger contexts and concurrent requests raise memory use. Keep sufficient free unified memory for the operating system.
What fits: GLM-5.3, GLM-5.3-Flash, Qwen3.8-Flash-Next, DeepSeek V4, Kimi K3, Inkling, and other large models. Flash-Next’s 125B backbone is only part of its roughly 180B-parameter checkpoint. Use the exact release’s serving recipe and account for all files.
Best for: Internal copilots, agent platforms, omnimodal services, governed deployment
Watch-outs: Some models can run with quantization and offload on smaller machines at a substantial speed cost. That is different from keeping the full model resident or serving multiple users.
File sizes here use decimal GB. They describe the named downloads checked on September 5, not total resident memory. Quantization, context length, concurrency, and offload can change both memory use and response time.
Hallucinations are only one part of the reliability story. Open models also fail through prompt sensitivity, poor tool arguments, visual grounding errors, license misunderstandings, and brittle long-context behavior.
The most common failures are fabricated facts, false confidence, stale knowledge, malformed JSON, and plausible-but-wrong code. Code models can also generate insecure or license-sensitive output.
Expect OCR misses, object misidentification, incorrect grounding, and overconfident descriptions of partially visible content.
The main problems are prompt drift, poor typography, inconsistent identity, and weak fine-grained control unless you add editing and conditioning tools.
The biggest issues remain temporal flicker, identity drift, motion incoherence, and long runtimes for short clips.
License fit is not cleanup work after the benchmark review. It is one of the first filters. Many teams waste time evaluating models they cannot legally or economically ship.
| License pattern | Best for | Examples | Watch-out |
|---|---|---|---|
| Apache 2.0 / MIT / OpenMDW | Commercial deployment and broad integration | Qwen3.8-27B, Muse Glimmer-30B, Gemma 4, EmbeddingGemma 2 (tagged Apache 2.0 in repo metadata; no LICENSE file), GLM-5.3-Flash, DeepSeek V4, gpt-oss, OLMo, Whisper, Qwen3-TTS, and Wan2.2. Read separate usage policies and third-party component terms. | Permissive weight terms do not establish full OSI open-source AI status. A derivative or quantization does not erase upstream obligations. |
| Llama 4 Community License + Acceptable Use Policy | Commercial use with strong ecosystem momentum | Llama 4 Scout, Llama 4 Maverick | Permissive for many uses, but not OSI-style open source. The 700M-MAU clause lives in the license; the restriction on multimodal use by EU-domiciled companies lives in the separate Acceptable Use Policy. Meta's newest open weights (Muse Glimmer) are Apache 2.0 instead. |
| Gemma terms / custom terms (legacy) | Practical use of older Gemma generations | Gemma 3 and earlier, including EmbeddingGemma 1 | Gemma 4 launched under Apache 2.0 on March 31, 2026, and Google's own Gemma Terms page now says "For Gemma 4 terms, see the Gemma 4 license." Gemma 3 and earlier remain under the custom terms. EmbeddingGemma 1 is manually gated under the Gemma terms; EmbeddingGemma 2 is ungated and tagged Apache 2.0. The family name alone is not a license signal. |
| OpenRAIL / Responsible AI licenses | Creative or research use where behavioral restrictions are acceptable | SDXL (CreativeML Open RAIL++-M), ControlNet weight distributions, BigCode OpenRAIL-M | Behavioral restrictions and downstream obligations can affect productization. |
| Custom commercial-use conditions | Case-specific commercial and research uses | Qwen3.8-Flash-Next: separate-license condition for MaaS or AI-work-assistant businesses without a revenue floor, with an internal-use exception. GLM-5.3: security review for MaaS operators above $10B aggregate licensee-and-affiliate revenue over twelve months. Kimi K3: a $20M aggregate-revenue trigger for MaaS operators, with exceptions. Qwen3.8-2.4T-A95B: a separate $50M aggregate-revenue condition. Mistral Medium 3.5: a $20M monthly-revenue condition. | These are distinct tests, not a universal revenue rule. Read business definitions, affiliate aggregation, attribution thresholds, and internal-use or hosted-provider exceptions in the linked licenses. |
| Non-commercial and territory-limited weight licenses | Research, experimentation, internal evaluation | Voxtral TTS (CC-BY-NC 4.0), FLUX.1-dev, FLUX.2 [klein] 9B, AudioCraft weights, and ProPainter. The custom SAM License has its own use restrictions. | Research or internal use is not automatically permitted by every restricted license. Check the intended activity and the actual files used. |
Treat every checkpoint as its own legal object. Do not assume the family name tells you the full commercial story.
Choosing a model without choosing a serving and evaluation stack is incomplete. The stack determines latency, batching, observability, and how painful future model swaps will be.
Best for: Fastest path to local testing on one GPU or a Mac
Strengths: A practical route to testing GGUF or other supported quantizations on local hardware. Match the runtime version to the model architecture, chat template, vision projector, and tool parser.
Limits: Context size, offload, and batching affect memory and latency. Do not assume every newly published architecture is supported by an installed desktop app.
Best for: High-throughput production inference for MoE models
Strengths: Batching and specialized model execution for server deployments. The new Qwen and GLM cards link to model-specific serving recipes.
Limits: Recipes may depend on recent releases or patches; pin the working version and benchmark your workload.
Best for: NVIDIA-centric optimized serving
Strengths: GPU-specific inference optimization for supported NVIDIA hardware and model architectures. Verify the current model support matrix before selecting a checkpoint.
Limits: More specialized setup and infra assumptions.
Best for: Custom workflows and research flexibility
Strengths: Best ecosystem for model experimentation, adapters, and editing pipelines.
Limits: Requires more assembly than end-user desktop tools.
Best for: Creative image and video workflows
Strengths: Visual pipeline building, strong community extensions, easy iteration. The standard path for Wan2.2, LTX-2.5 and open-video generation workflows.
Limits: Operational governance is weaker than code-first stacks.
Best for: Agents, tool use, and workflow orchestration
Strengths: Useful abstractions for state, retrieval, and multi-step execution.
Limits: They do not fix weak evals or poor model choices for you.
Use these as default launch points, not as permanent architecture decisions.
Test Qwen3.8-27B or Muse Glimmer-30B on the hardware you have, with a modest context budget and a small set of representative tasks. Use a smaller Gemma 4 or gpt-oss-20b configuration when memory is tight.
Compare Qwen3.8-27B with Qwen3-Coder-Next on your repositories. Coder-Next’s 48.41 GB Q4_K_M artifact needs runtime headroom beyond the download size; use a larger-memory machine or explicit offload.
Use SDXL or an appropriately licensed FLUX checkpoint for image experiments. For video, compare the documented Wan2.2 or LTX-2.5 workflow with your footage and hardware. Confirm output quality, processing time, and current weight terms before production use.
Compare MiniCPM-o 4.5 and Qwen3-Omni-30B-A3B-Instruct for audio-visual interaction, Whisper or Qwen3-ASR for transcription, and Qwen3-TTS or Chatterbox for synthesis. Check whether each variant produces speech or only text; benchmark end-to-end latency and review voice permissions.
Restrict tool permissions, sandbox execution, and evaluate prompt injection with representative untrusted inputs. A guard model is one layer. Meta’s AgentDojo result describes one benchmark setup and is not a measured baseline for every agent.
Prioritize license clarity (per-checkpoint, not per-family), eval discipline, and serving fit over raw leaderboard rank. vLLM-class serving plus Qwen3.8-27B, Muse Glimmer-30B or DeepSeek-V4-Flash-0731 (MIT) is usually the right first step. If a checkpoint's license carries a revenue trigger (Kimi K3, Qwen3.8-2.4T-A95B, Mistral Medium 3.5), have counsel read it before the evaluation starts.
Pick one workflow, one evaluation harness, one hardware target, and three candidate models. Anything broader becomes expensive research theater.
This page is built from model cards, technical reports, official repositories, standards bodies, and tooling documentation. The goal is practical decision support, not hype-driven ranking.