MiniMax Releases M3.1 Flash Open Weights: a 428B MoE in a 252 GB Checkpoint

TL;DR

On October 23, MiniMax published the weights for MiniMax-M3.1-Flash on Hugging Face, four days after the model went live on its API. It is a native multimodal mixture-of-experts model with about 428B total parameters and 23B active per token, a 1M-token context window, and text, image and video input. The parameter count matches MiniMax-M3. The difference is that M3.1 Flash was trained with quantization in the loop, so the checkpoint is 252 GB against 854 GB for M3 in BF16. MiniMax charges $0.15 per million input tokens and $0.60 per million output tokens for it. On MiniMax's own benchmark table it leads Opus 5 and GPT-5.6 Sol on agentic automation and trails both on harder coding and reasoning tests. The license is open-weight with conditions: commercial users must credit MiniMax and notify it, and companies above $20M in yearly revenue need written permission.

What shipped

MiniMax announced M3.1 Flash on October 19 through its API, M Plan and MiniMax Code, and said weights would come "later." They arrived on Hugging Face on October 23 as a 72-file repo, with serving recipes ready the same day. Together AI is hosting it too, so you can call it without owning the hardware.

The model card lists three changes from M3:

• Sparse attention in every layer. M3 kept three full-attention layers. M3.1 Flash replaces them with MiniMax Sparse Attention (MSA), so every layer is now sparse.
• Quantization-aware training. The routed experts and the attention paths were quantized during training. According to the vLLM recipe, dense layers ship in MXFP8 and experts in NVFP4.
• DSpark speculative decoding. A draft head sits in the repo's dspark/ directory. It replaces the EAGLE-like multi-token prediction M3 used, so you don't need a separate draft model.

Quantization-aware training is the change that matters most for you. Normally a lab trains in high precision, releases BF16, and the community compresses it afterward, and the compression costs some quality. With QAT the model learns its weights while already living in 4-bit experts. Post-hoc quantization is like taking in a finished suit. QAT is cutting the cloth to the smaller size from the start, so the 4-bit version is the official model and not a lossy copy of it. The card is careful on one point: "Results are for the official weights; third-party quantizations may differ."

What it takes to run

252 GB is a lot smaller than 854 GB, but this is still a server model. The vLLM recipe needs vLLM 0.33.0 or later and lists three hardware setups:

• NVIDIA Hopper (H100 or H200): a full 8-GPU node at tensor parallel 8, with FP8 KV cache. The example caps context at 262,144 tokens, and DSpark is not enabled on Hopper.
• NVIDIA Blackwell (GB200, GB300, B200, B300): tensor parallel 4, with an NVFP4 KV cache and DSpark enabled (the example config drafts three tokens).
• AMD MI355X: tensor parallel 4. vLLM re-quantizes the NVFP4 experts to MXFP8 when it loads them.

SGLang and TokenSpeed have recipes as well. MiniMax says Transformers, KTransformers and unsloth support is "in progress." If you were hoping for a pair of gaming cards, the weights alone fill 252 GB before you allocate a single byte of KV cache.

The benchmarks, read carefully

All of these numbers come from MiniMax's table. Each benchmark has its own harness note, and several of them ran through Claude Code at different context lengths. Treat it as a vendor scorecard until independent leaderboards catch up.

Agentic automation is where it wins. On Automation-Bench v1.0.6 it scores 62.8, ahead of DeepSeek V4.1 Flash-1017 at 60.5, Opus 5 at 50.3 and GPT-5.6 Sol at 45.8. It ties Opus 5 at 76.2 on BankerToolBench and edges GPT-5.6 Sol on LOCA-Bench 256K, 60.3 to 60.0.

It loses on harder engineering. On Terminal Bench 4.0 it gets 38.1 to Opus 5's 51.8. On NL2Repo-Bench it scores 61.2 to Opus 5's 75.3. On text-only HLE without tools it gets 43.4 to Opus 5's 55.6. On OSWorld 2.0 computer use it scores 61.7, below GPT-5.6 Sol at 65.7 and Opus 5 at 70.6, though MiniMax says that is the top score among Chinese models. On DeepSWE v1.1 it scores 73.7, level with the closed models (Opus 5 74.0, GPT-5.6 Sol 73.0) and trails DeepSeek's October Flash at 77.0.

The open-weight comparison is closer. M3.1 Flash beats Kimi K3 and GLM-5.3 on most rows where both have scores, but GLM-5.3 leads on CyberGym (84.5 to 82.2) and Kimi K3 leads on CharXiv reasoning (84.8 to 82.5). DeepSeek V4.1 Flash-1017 beats it on Terminal Bench, DeepSWE, NL2Repo and BabyVision.

Price and speed

MiniMax's API price is $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. OpenRouter lists the same $0.15/$0.60. OpenRouter's M3 listing is $0.30/$1.20, so on that marketplace the new model costs half as much per token.

The launch post also claims online inference is "nearly three times faster," output runs at 200+ tokens per second with a P50 of 170+, and agentic usage costs 70% less than the predecessor. MiniMax doesn't name the predecessor or publish a method for these numbers, so they are marketing until someone reproduces them on their own hardware.

The same post says an early version of the model ran anonymously as "Space Bunny" and handled more than 200 trillion tokens before anyone knew whose it was. Stealth launches are the new soft open, and apparently the bunny had a heavy workload.

The license is the catch

The weights are downloadable, but this is not an MIT release. The MiniMax Community License grants free use for non-commercial purposes. For commercial use:

• You must "prominently display" Built with MiniMax M3.1-Flash on a website, UI, blog post, about page or product docs.
• Below $20M in yearly revenue, you send MiniMax a one-time notice email.
• Above $20M, you need prior written authorization.
• Commercial use explicitly covers hosted APIs and fine-tuned or otherwise post-trained derivatives.
• An appendix bans uses including military purposes and harmful disinformation.

These are the same terms M3 shipped under, so nothing got worse. For most small teams it amounts to an email and a credit line. For a company above the revenue line, it is a negotiation.

Key Takeaways

• MiniMax-M3.1-Flash is now open-weight on Hugging Face: about 428B total and 23B active parameters, 1M context, text, image and video input.
• Quantization-aware training puts the experts in NVFP4 and the dense layers in MXFP8, so the official checkpoint is 252 GB against 854 GB for M3 in BF16.
• It still needs a full 8-GPU H100/H200 node, or four Blackwell or MI355X GPUs, with vLLM 0.33.0+.
• MiniMax's own table has it leading Opus 5 and GPT-5.6 Sol on Automation-Bench (62.8) and trailing them on Terminal Bench 4.0, OSWorld 2.0 and HLE.
• API pricing is $0.15/$0.60 per million tokens, half of M3's OpenRouter rate. The speed and 70% savings claims are unverified.
• The Community License requires credit and a notice for commercial use, and written permission above $20M in yearly revenue.

Sources: MiniMax-M3.1-Flash model card (Hugging Face), MiniMax Community License, MiniMax launch post (Oct 19, 2026), vLLM recipe for MiniMax-M3.1-Flash, Together AI model page, OpenRouter listing, MiniMax-M3 model card, MiniMax Sparse Attention (arXiv)