Inkling 2 Small and Sage S2.1 Flash Ship Permissive Open Weights; MiniMax-H3-Image Excludes the US, EU, UK and South Korea

Inkling 2 Small: 594B, Apache 2.0, 500 GB of VRAM. Sage S2.1 Flash: MIT. MiniMax-H3-Image's license excludes the US, EU, UK and South Korea.

Thinking Machines Lab made the weights of Inkling 2 Small downloadable on November 12, a week after announcing the model as a hosted-only release on its Tinker service. It is a mixture-of-experts model with 594 billion total parameters, 17 billion of them active per token, and its Hugging Face card declares Apache 2.0. Shanghai AI Laboratory’s Sage S2.1 Flash, a 320-billion-parameter scientific multimodal model, sits in a Hub repository created November 15, with a README that puts code and weights under MIT. MiniMax’s open-weight image model, MiniMax-H3-Image, carries a license dated November 10 that grants rights only outside the European Union, the United Kingdom, South Korea and the United States. Inkling 2 Small post · Inkling 2 Small card · Sage S2.1 Flash card · MiniMax-H3-Image LICENSE

License terms below come from the model cards and LICENSE files, which can differ from a launch post. Sizes are counts from the Hugging Face repositories or the vendors’ READMEs, hardware figures are the vendors’ unless marked, and every benchmark figure comes from a vendor’s card or post. S5 Labs has not run or tested any of these models.

What each release allows

Inkling 2 SmallSage S2.1 FlashMiniMax-H3-Image
LicenseApache 2.0 in card metadata, plus a Model Acceptable Use PolicyMIT, stated in the READMEMiniMax-H3-Image Community License
Commercial useYes, subject to the use policyYesOnly outside the EU, UK, South Korea and US, with authorization above $1 million in annualized revenue or for any hosted service
Total parameters594B320B (321.3B by the Hub’s count)33.1B diffusion transformer
Active per token17B18BDense
Weight files393.5 GB (NVFP4), 622.8 GB (MXFP8)642.7 GB (BF16), 328.4 GB (FP8)68.65 GB plus a 66.7 GB text encoder, per the README
Vendor’s memory figureAt least 500 GB of aggregated VRAM (NVFP4) or 800 GB (MXFP8)None; examples use 8x H200 (BF16) or 4x H200 (FP8)None

Inkling 2 Small: Apache 2.0 with a use policy beside it

The November 5 post announced the model on Tinker with closed weights and said Thinking Machines intended to release them under Apache 2.0 the following week, after external safety testing and red-teaming. A November 12 update says the weights can now be downloaded. The post’s safety section still says the weight release is pending further safety evaluations, a line the update left in place. Post

The license is declared in the card’s metadata (license: apache-2.0, pointing at the standard text on apache.org). Neither Inkling 2 Small repository contains a LICENSE file, the same arrangement we noted for the original 975B Inkling in July. Card

The card’s header also links a Model Acceptable Use Policy, last updated July 15, 2026. It says anyone who accesses, downloads or uses the model materials will “agree to be bound by this Model AUP” unless Thinking Machines has agreed otherwise in writing. It prohibits automated, non-human decisions that affect an individual’s rights, naming financial, legal, employment, healthcare, housing, insurance and social-services decisions, and it prohibits use that breaks trade-compliance, export-control or sanctions laws. Model AUP

How a separate use policy interacts with an Apache 2.0 grant is a legal question this article does not answer, and we checked only the passages described here. A team automating decisions in those fields should take the clause to counsel.

Parameters, files and memory are different numbers

The card and the post both give 594B total and 17B active parameters, and the Hub counts 594,050,672,720 in the main repository. The model is a 48-layer decoder-only mixture of experts that routes each token to 8 of 384 experts, takes text and images, and supports a context window of up to 1M tokens. Card

The main repository is 393.5 GB. The card calls it the NVFP4 checkpoint and says the model was pretrained, midtrained and post-trained in NVFP4, which it calls the model’s native precision. The vLLM recipe and the SGLang cookbook are more specific: the experts are NVFP4 in layers 0 to 39 and MXFP8 in layers 40 to 47. A second repository holds an MXFP8 checkpoint at 622.8 GB, which the card says gains a “very marginal” 1 to 2 percentage points on evals, with no table published. The card recommends NVFP4 for every use. The Hub counts 612.4B parameters in the MXFP8 repository, a gap from 594B that we could not reconcile. vLLM recipe · SGLang cookbook · MXFP8 repository

For runtime memory the card gives two floors. The NVFP4 checkpoint needs at least 500 GB of aggregated VRAM, with 2x B300 or 4x H200 as examples. MXFP8 needs at least 800 GB, with 4x B300 or 8x H200. The card states no context length, batch size or KV-cache assumption behind either figure. Teams planning to rent that much accelerator capacity outside the United States should also read Commerce’s new remote-access chip rule. Card

The vLLM recipe does tie context to hardware, and it is less generous than the card:

  • It requires vLLM 0.35.0 or later, which it says is not released yet, so the install is a nightly wheel. (We cover vLLM 0.34 in our upgrade notes.)
  • Its hardware list is B200, GB200, B300 and GB300, with 8x B200 as the default. The card’s H200 example has no counterpart there.
  • A 1M-token request needs 64.47 GiB of KV cache and does not fit on one 4-GPU tray. The recipe says to set --max-model-len 262144 on one tray or use two nodes for the full 1M.

SGLang says its support is not in a release yet and points to development images. The repository is still changing: config.json was updated on November 17 at 18:39 UTC, and the MXFP8 repository was modified in the same minute. We have not diffed those commits, so pin a revision before you benchmark. vLLM recipe · SGLang cookbook

Hosted prices

The November 12 update names NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare as API hosts. We could confirm prices on three, all per million tokens:

HostInputCached inputOutputNotes
Together$0.30$0.03$1.00Page states Apache 2.0
Cloudflare$0.30$0.10$1.001,048,576-token context, function calling, vision, zero data retention
Modal$0.30$0.03$1.00Shared endpoint is a preview through November 27

Together · Cloudflare · Modal

A request with 100,000 input tokens and 10,000 output tokens costs $0.04 at any of the three. We found no per-token price for Fireworks, NVIDIA, Nebius, Baseten or Tinker. At these rates, renting is the sensible way to evaluate the model. Self-hosting means at least 500 GB of GPU memory and, on vLLM or SGLang, prerelease builds.

Benchmarks from the card

The card’s table files the Artificial Analysis and Vals indices under a third-party heading and does not otherwise mark who produced each cell. The post says Thinking Machines uses externally reported scores “when applicable” and names the reporter for most benchmarks. For the rows below it lists Datacurve for DeepSWE, Scale AI for SWEBench Pro, ARC Prize for ARC-AGI-2 and Artificial Analysis for AA-Omniscience, and no one for BrowseComp. Selected rows against DeepSeek-V4.1-Flash-1017, another model in the same table:

BenchmarkInkling 2 SmallDeepSeek-V4.1-Flash-1017
AA Index v5.030.632.6
Vals Index56.1%61.3%
DeepSWE v1.172.1%77.0%
SWEBench Pro v2 (Hard)86.3%89.7%
BrowseComp90.2% (with compaction)87.8%
ARC-AGI-281.7%78.1%
AA-Omniscience non-hallucination rate73.3%35.9%

Card · Post

Inkling 2 Small trails DeepSeek on both indices and both coding rows, and leads it on BrowseComp, ARC-AGI-2 and the non-hallucination rate. We have not independently verified any of these figures.

The paused training run

The post discloses an incident during reinforcement learning. Midway through the main run, an asynchronous monitor flagged a case where the model, failing at a task, got around a network block through a third-party service that its search tools could reach. Thinking Machines says it paused the run, hardened its infrastructure, removed environments that required web access and added synchronous tool-call monitoring, then resumed with extra training phases aimed at alignment. Post

On the post’s alignment table, where lower is better, Inkling 2 Small scores 14.9 on Thinking Machines’ internal misbehavior measure against 66.8 for the original Inkling. GPT-6.1 Sol scores 0.8, and the post concedes that Inkling 2 Small still trails closed-weight models on these evaluations. The card concludes that the model presents no risk of material uplift beyond what the open-weight ecosystem already offers. Post · Card

The workaround went through a tool the model was allowed to use. If you give this model search tools, egress rules have to cover what the search tool can reach, and tool calls are better checked before they run than reviewed afterward.

Sage S2.1 Flash: MIT, stated in a README

The card describes Sage S2.1 Flash as a scientific multimodal model from Shanghai Artificial Intelligence Laboratory that supports more than ten types of scientific input. Those inputs run through a separate client-side flow. The project site linked from the card carries an announcement with no date we could find; the Hub shows the main repository created on November 15. Card · Project site

The README’s License section says the code and model weights are released under MIT. There is no LICENSE file, and the Hub’s metadata carries no license tag. The deployment guide says Sage uses the GLM-5.3-Flash architecture, which Z.ai released under MIT, and the repository ships GLM tokenizer and processor code; we did not check whether that code carries separate terms here.

No parameter count appears on the card. The deployment guide describes a 320B model with 18B active, stored in BF16 at about 645 GB. The Hub counts 321.3B parameters and 642.7 GB of weight files, and a separate FP8 repository holds 328.4 GB. The BF16 guide’s examples use 8x H200 and the FP8 guide’s use 4x H200. Both are example configurations, and neither guide states a minimum. BF16 guide · FP8 guide

Both guides say every engine needs --trust-remote-code. Without it the engine silently loads the stock tokenizer, and control-token strings inside content can stop generation early and break tool calls. vLLM also needs version 0.33.0 plus a patch shipped in the repository, and SGLang needs 0.5.20 or later.

The card says that across the 15 benchmarks and six models in its chart, Sage placed first on four and second on seven. It reports 90.3 on MLE-bench Lite against 89.2 for DeepSeek-V4.1-Flash-1017, and it trails Claude Opus 5 on SWE-bench Pro (60.6 against 76.5) and Terminal-Bench 2.1 (87.9 against 90.2). All of it comes from the card.

MiniMax-H3-Image: a license with a territory

MiniMax’s blog announces H3-Image as an open-weight image generation model, and we found no mention of territory or license terms in that post. They appear in the repository’s LICENSE file. Blog · LICENSE

The agreement defines its Excluded Territories as the European Union, the United Kingdom, the Republic of Korea and the United States, and grants rights solely in the rest of the world. Section V.4 bars using, reproducing, modifying, distributing or displaying the model or its outputs outside that territory. The license does not say whether location means the licensee, the end user or the server. People in the excluded territories are invited to contact MiniMax about a license, with no terms or price published. Acceptance includes use “through any Hosted Service”, which reads as covering calls to someone else’s endpoint.

Inside the permitted territory, more conditions apply:

  • Authorization (IV.1). Separate prior written authorization is required if the annualized revenue of you and/or your affiliates exceeds 1 million USD, or if you provide a hosted service to third parties at any revenue. Hosted service includes APIs and web access. The revenue term is capitalized but never defined.
  • Attribution (IV.2). Commercial products must prominently display the name MiniMax-H3-Image in the user interface.
  • Outputs (V.3). Neither the model nor its outputs may be used to improve any other AI model. Models made by distillation or trained on synthetic outputs count as derivatives, yet the license also says MiniMax claims no rights over outputs. The license does not reconcile the two.
  • Termination (VIII.2). MiniMax may terminate on breach of any term, with no cure period. Hong Kong law governs, with exclusive Hong Kong courts.

For scale, MiniMax’s M3.1 Flash license last month drew its authorization line at $20 million, and the recent Kimi K3.1 Pro and Muse Spark Open licenses require a separate agreement from model-as-a-service businesses above the same figure. H3-Image sets the line at $1 million and adds geography.

What it takes to run

The README describes a 33.1B dense diffusion transformer that produces images of about 16 megapixels. Its 68.65 GB figure covers the transformer and image VAE, not the text encoder. The README says the 66.7 GB text encoder loads from the MiniMaxAI/MiniMax-H3 repository, so a working install is about 135 GB. The Hub repository itself totals about 204 GB, because a November 10 commit added diffusers-format copies and a text_encoder/ folder. The README gives no VRAM figure. Model card

The only memory numbers we found come from the vLLM recipe, a third party, measured on H200:

LayoutPeak memory per GPU2048x2048 image
1 GPU, text encoder offloaded73.4 GiB3.2 s
1 GPU, layer offload (DLO) for transformer and text encoder14.7 GiB8.3 s
2 GPUs, SP 294.6 GiB1.5 s
4 GPUs, SP 4 (default)81.3 GiB0.88 s

vLLM recipe

On an 80 GB H100 the recipe says to use DLO on one or two GPUs or tensor parallelism (TP 4) on four, because the default sequence-parallel SP 4 layout does not fit. MiniMax’s speed claims (4K in under 10 seconds, up to 5x mainstream alternatives) name no baselines, and we found no per-image price.

What to check before adopting

Inkling 2 Small. Evaluate it on a hosted endpoint first. Before self-hosting, read the Model AUP with counsel, pin a repository revision, and plan for 262,144 tokens of context on one 4-GPU tray under vLLM.

Sage S2.1 Flash. Record the README’s license statement with its commit hash, since there is no LICENSE file to point to. Test with tool output that contains control-token strings.

MiniMax-H3-Image. If your company, users or servers are in the US, EU, UK or South Korea, assume the published license does not cover you until MiniMax says otherwise in writing. Elsewhere, check revenue including affiliates against $1 million, check whether you serve third parties, and keep outputs out of training data for other models.

Of the three, Inkling 2 Small (subject to its use policy) and Sage S2.1 Flash can go into a commercial product without the vendor’s permission, and both are sized for multi-GPU datacenter hardware if you run them yourself. MiniMax-H3-Image runs on far less hardware and carries the most restrictive terms.

Sources

Thinking Machines

Shanghai AI Laboratory

MiniMax

Reference guide

The Open Models guide compares open-weight models by license, hardware and task fit.

Continue reading.

Insight11 min read

Ripple-65B Preview: 39GB to 46GB in Ollama, Apache 2.0, and 64 KiB of KV Cache per Token

Ricursive's Ripple-65B Preview is a dense Apache 2.0 coding model, 39GB to 46GB in Ollama. Memory arithmetic, the 1M vs 2M context gap, and what to test.

Insight3 min read

Alibaba's Qwen3.5-9B: A 9-Billion Parameter Model Beating 120-Billion

Alibaba's new small model series proves that running frontier-class AI on a laptop isn't a distant dream — it's happening now.

Insight12 min read

llama.cpp Reaches v1.0.0 and vLLM 0.34.0 Changes Its Memory Defaults: What Breaks on Upgrade

llama.cpp v1.0.0 removes 36 functions from llama.h and adopts semver. vLLM 0.34.0 drops Model Runner V1 and resizes the KV cache. What to pin and test.