MiniMax announced M3.1 Flash on October 19, an agent-focused model priced at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, available through the MiniMax API, M Plan and MiniMax Code. MiniMax blog Pricing The same blog post says “Model weights will be released later.” Yet a public repository named MiniMaxAI/MiniMax-M3.1-Flash is already on Hugging Face, created on October 17 and last modified on October 23, with 49 safetensors shards and a LICENSE file. Hugging Face API The two routes to the model are two different decisions. The API is a per-task cost question. The weights are a licensing question, and a hardware question that the published size figures do not yet settle.
S5 Labs has not run the model. Every benchmark, speed and cost figure below is MiniMax’s unless marked as our arithmetic.
The API: what a token costs
MiniMax’s pay-as-you-go page lists three prices for M3.1 Flash and no tiers. It does not list a cache-write price or a context limit for the model. Pricing The older M3 is on the same page, marked “Permanent 50% off”, with struck-through list prices and a Priority tier at 1.5 times Standard (service_tier=priority). Pricing
| Model and tier (per 1M tokens) | Input | Output | Cache read |
|---|---|---|---|
| M3.1 Flash | $0.15 | $0.60 | $0.015 |
| M3 Standard, up to 512k input (current) | $0.30 | $1.20 | $0.06 |
| M3 Standard, up to 512k input (struck-through list) | $0.60 | $2.40 | $0.12 |
| M3 Standard, over 512k input (current) | $0.60 | $2.40 | $0.12 |
| M3 Priority, up to 512k input | $0.45 | $1.80 | $0.09 |
| M3 Priority, over 512k input | $0.90 | $3.60 | $0.18 |
Source: MiniMax pricing page, checked October 23.
The blog’s headline cost claim is that “for agentic tasks, the model’s overall usage cost is 70% lower than its predecessor.” MiniMax blog The predecessor is not named and no method is given. Our arithmetic: against M3’s current discounted rates, M3.1 Flash is 50% cheaper per token on both input and output; against M3’s struck-through list rates, it is 75% cheaper (for M3 inputs up to 512k tokens). A 70% figure for whole tasks could come from fewer tokens per task, a different baseline, or a mix, and the blog does not say which.
On speed, MiniMax says output “can reach 200+ TPS” and that median (P50) speeds are “consistently reaching 170+ TPS”. MiniMax blog Those are the vendor’s numbers for its own service, with no test conditions attached.
What MiniMax says it scores
The model card compares M3.1 Flash on agent benchmarks against DeepSeek V4.1 Flash (two versions), Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5, and the card does not name the harness or settings. The blog carries the headline scores in its text, with its comparison data in chart images we could not read. MiniMax blog Model card A selection of the card’s M3.1 Flash column:
| Benchmark (MiniMax’s runs) | M3.1 Flash |
|---|---|
| DeepSWE v1.1 | 73.7 |
| OSWorld 2.0 | 61.7 |
| Terminal Bench 4.0 | 38.1 |
| Agents’ Last Exam | 29.6 |
| CyberGym | 82.2 |
| BankerToolBench | 76.2 |
| LOCA-Bench 256K | 60.3 |
| NL2Repo-Bench | 61.2 |
| SWE Refactor Bench | 23.0 |
| HLE (without tools) | 43.4 |
| BabyVision (with tools) | 91.0 |
Source: model card README.
The card’s table shows BankerToolBench at 76.2 for both M3.1 Flash and Opus 5. Model card The blog’s labels and rounding differ a little from the card’s: it calls the 61.7 “OSWorld V2” where the card says “OSWorld 2.0”, and gives CyberGym as “82%” where the card says 82.2. MiniMax blog The card also notes its results are “for the official weights; third-party quantizations may differ.” Model card Independent runs of these numbers are not covered here.
The weights: a live repo and a blog that says “later”
The Hugging Face repository is public and not gated. Its license metadata reads “minimax-community” with the tag “other”. The 49 shards are 48 main-model files plus one dspark/ draft-model shard with its own config. Hugging Face API The discussions tab has recent posts celebrating the release and one report that vLLM does not split reasoning from the answer when streaming, with no reply from MiniMax shown. Discussions MiniMax’s own release-notes page describes the October 19 entry as “an ultra-efficient open-weight model for everyday work.” Release notes
So the blog says later, the release notes say open-weight, and the files are downloadable. MiniMax has not published anything that reconciles those. The repo sits under the MiniMaxAI organization, which is the only basis for calling it official; the README does not link to the blog. If you plan to pull the weights, confirm they are still there and unchanged since our October 23 check (the current commit is 54e4e69).
Size figures disagree, so hold off on hardware planning
The README describes a native multimodal mixture-of-experts model (text, image and video inputs, text output) with “approximately 428B total; 23B activated per token” and a context window of “up to 1M tokens”. Model card The config.json matches on shape: hidden size 6,144, 60 layers, 128 local experts with 4 routed per token, and max_position_embeddings of 1,048,576, all under text_config. config.json
The shipped weights are quantized. The config’s quantization_config sets quant_method to mxfp8 with dynamic activations, and the MoE experts to NVFP4 in the nvfp4-pack-quantized format; the MoE gates, lm_head, embeddings, vision tower and multimodal projector are listed as ignored layers. config.json The README says the routed experts and attention paths were quantized during training (QAT). Model card
This is why the parameter count looks wrong at first glance. Hugging Face’s safetensors metadata reports 220,490,854,016 entries: 206.6 billion U8, 10.6 billion F8_E4M3, 3.3 billion BF16 and 46 million F32. Hugging Face API Our arithmetic: if each U8 entry packs two 4-bit values, the expert weights hold about 413 billion parameters, and adding the other tensors gives roughly 427 billion, in line with the card’s 428B. MiniMax has not stated this reconciliation, and it does not give an on-disk size or memory requirement anywhere. Do not read 220B as the model’s size, and do not budget hardware from either number until someone publishes a measured download size and a serving footprint at a stated context length.
Two other README details need a close read. On attention, the “What’s New” section says the first three full-attention layers are replaced with sparse attention, while “Model at a Glance” lists MiniMax Sparse Attention “throughout the model”; the two are compatible but the card does not spell out how many layers remain full attention. Model card On decoding, the card says DSpark speculative decoding “replaces the earlier EAGLE-like MTP approach” and that the draft head ships in the repo. Model card
For serving, the README recommends SGLang, vLLM and TokenSpeed, with Transformers, KTransformers, unsloth and ATOM in progress; sampling is temperature 1.0 and top_p 0.95, and reasoning effort (low to max) is set in the chat template. Model card
The license: a notice below $20 million, authorization above
The repository’s LICENSE file is the MiniMax Community License. S5 Labs read the raw LICENSE file on October 23, and the terms below follow its text. It is not a lawyer’s reading, so have counsel check it before relying on it. LICENSE
Commercial use is defined as use primarily intended for commercial advantage or monetary compensation, and it covers paid products or services built on the model, commercial use through hosted or cloud APIs, and commercial deployment of fine-tuned or post-trained versions. Commercial users must prominently display “Built with MiniMax M3.1-Flash” on a related website, UI, blog post, about page or product documentation. They must also contact [email protected], and which message depends on revenue. If the products or services generate more than 20 million US dollars (or the equivalent in other currencies) in yearly revenue, the user needs separate, prior written authorization, requested with the subject line “M3.1-Flash licensing - authorization request”. Otherwise a one-time notice is enough, with the subject “M3.1-Flash licensing — notice”. Non-commercial use is granted free of charge, and the copyright and permission notice must stay in all copies. The appendix of prohibited uses covers content that breaks the law, any military purpose, exploiting or harming minors, false or misleading information intended to cause harm, and promoting discrimination or hate based on protected or systemically marginalized characteristics. LICENSE
The revenue test is worded around “such products and services”, meaning the commercial products or services that use the model or its derivatives. The text does not mention the licensee as a whole or its affiliates, so on its face the threshold is measured per product or service rather than per company, though the file does not say how revenue is allocated when a company runs several. Liquid’s LFM license, by contrast, counts revenue across the whole legal entity and its affiliates, and we read that one from the raw file. The file has no termination clause and no governing-law or venue clause; its only other term is a standard “as is” warranty disclaimer.
Against the other recent open-weight agent models, this sits in the middle. Qwen3.8-Max put its gate on hosted-service businesses at $50 million. GLM 5.3 Flash is MIT, and Tencent’s Hy4 Preview is Apache 2.0, with the hardware bill as the catch. At the small end, Recursive’s Flint is an Apache 2.0 coding model with a 23 GB local build. MiniMax asks for an attribution line and an email from every commercial user, which is more paperwork than Apache and less than a hard revenue cutoff.
Two decisions, checked separately
For the API, the questions are ordinary. Run your agent’s real traces and compare total tokens per completed task against whatever you use now, since the 70% claim’s baseline is undisclosed. Confirm the context limit and maximum output with MiniMax before relying on the card’s million-token figure; the pricing page does not state either. Ask about cache-write charges, which are not listed.
For the weights, start with the license file, read raw, and get counsel’s read on the per-product revenue test before anyone fine-tunes. Then wait for a measured size. The quantization is baked in, the serving stacks are SGLang, vLLM and TokenSpeed today, and one open discussion already reports that vLLM does not split reasoning from the answer when streaming. The open-source model guide tracks which of these models have measured deployment footprints.
What is still unknown
MiniMax has not explained why the blog says the weights come later while the repository is live. There is no stated download size, memory requirement, API context length or maximum output. The 70% cost claim has no named baseline. We have not checked for independent reproductions of the benchmarks, and the license leaves open how revenue is counted for a company with several products. None of these can be answered from the launch materials alone.