# MiniMax-M3.1-Flash (slim)

> MiniMax-M3.1-Flash is MiniMax's open-weight mixture-of-experts model for coding, agent work, computer use and cybersecurity, about 428B parameters with 23B active, taking text, images and video with up to 1M tokens of context.

- Full: https://www.anchorterminal.com/tools/minimax-m3-1-flash.md (~12,950 tokens) · this version ~2,530 tokens · JSON https://www.anchorterminal.com/tools/minimax-m3-1-flash.json · canonical https://www.anchorterminal.com/tools/minimax-m3-1-flash
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-25 · pricing https://www.anchorterminal.com/tools/minimax-m3-1-flash#pricing · reviews https://www.anchorterminal.com/tools/minimax-m3-1-flash#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 50.5/100 · rank #99 of 110 open-weight models · #77 in Open-weight models · not agent-ready · confidence high**


Assessment: A 428B mixture-of-experts model, 23B active, for agent work, coding and computer use, with image and video input, about 1M tokens of context and parallel tool calls. It needs a 4 to 8 GPU server for 250 GB of NVFP4 and MXFP8 weights, and vLLM and SGLang support was pre-release on 24 October 2026. Commercial use needs attribution and notice.

## Facts

- Kind: Open-weight model · vendor: MiniMax · category: Open-weight models · legal entity: MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore) · provenance 55/100
- Packages: huggingface `MiniMaxAI/MiniMax-M3.1-Flash`
- Auth: None · pricing: Free · x402: no · licence: MiniMax Community License (custom; commercial use needs a "Built with MiniMax M3.1-Flash" line and a one-time notice to MiniMax, and MiniMax's written authorisation above $20M yearly revenue)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash, read from the model card, the files and their metadata, the licence, the chat template, the configs, the commit history, MiniMax's launch post and release notes, and the runtimes' own code, recipes and libraries on 24 October 2026. Nothing was downloaded or run. MiniMax's hosted API is its own listing (minimax-api)
- Model: About 428B parameters with 23B active per token, per the card (vLLM's recipe counts 427.05B in the safetensors headers; the Hub counts 220.5B because the NVFP4 experts are packed two to a byte). MiniMaxM3SparseForConditionalGeneration, 60 layers, 128 routed experts with 4 per token plus one shared expert, MiniMax Sparse Attention in every layer, and a 32-layer vision tower. Text, image and video in, text out
- Files: 48 safetensors shards (250.12 GB), routed experts in NVFP4 and the rest in MXFP8, BF16 and F32, plus a DSpark draft (2.30 GB) in dspark/. No pickle files. Hugging Face lists a SHA-256 for every LFS file. The repository also holds Python configuration and processor code (configuration_minimax_m3_vl.py, processing_minimax.py, image_processor.py, video_processor.py) named in auto_map, and is tagged custom_code
- Licence: MiniMax Community License (Hub tag other), the same text as MiniMax-M3's with the model name changed. The grant covers non-commercial use, and section 2 allows commercial use if the product shows "Built with MiniMax M3.1-Flash" and MiniMax gets a one-time notice, or, above $20M of yearly revenue, MiniMax's prior written authorisation. Commercial use includes paid products, commercial APIs and serving fine-tuned derivatives. A Prohibited Uses appendix bans military use among others. No change or revocation clause
- Runtimes: llama.cpp from release b11861 (24 October 2026), which has the minimax-m3 architecture with MSA since July and the NVFP4 scale fix for M3.1 Flash; no publisher or library GGUF. vLLM (recipe min version 0.33.0, nightly image only; released v0.32.0 has the architecture but the recipe says it isn't enough) and SGLang (unmerged pull request or dev image) per their own docs. Transformers v5.20.2 has minimax_m3_vl, but the card says Transformers support is still in progress. Ollama is cloud only and LM Studio has no catalogue page
- Context: 1,048,576 tokens in config.json with no rope scaling, and up to 1M on the card. tokenizer_config.json says 40,960,000 as model_max_length. vLLM's Hopper command caps it at 262,144
- Tools and reasoning: The chat template declares tools as JSON inside tools tags and asks for calls as invoke blocks with XML-expanded parameters inside a minimax-namespaced tool_call block, several per block. Tool results go in a tool role inside response tags. The card says nothing about tools; the parsers are minimax_m3 in vLLM and minimax-m3 in SGLang. Thinking is adaptive, reasoning comes back inside mm:think tags, and reasoning_effort (low, medium, high, xhigh, max) is set per request through the template. Separate root (higher priority) and system or developer (lower priority) roles. Recommended sampling temperature 1.0, top_p 0.95
- Hardware: The card states none. vLLM's recipe runs it on 4 GB200, GB300, B200, B300 or MI355X GPUs, or 8 H100 or H200 GPUs without DSpark or video. Nothing fits a 128 GB machine
- Training data: Not described. The card mentions quantisation-aware training only. The template's default system prompt gives a knowledge cut-off of January 2026

- Scores: Capability 64, Reach 40, Risk 35, Agent interface 70, Documentation and upkeep 41 · total over the 7 assessed categories
- Why: Capability, From 6 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns nothing. · Risk, File safety earns 15, safetensors only, less 10 because the SGLang route the card links passes trust_remote_code for the repository's config… · Agent interface, The chat template earns 20 for chat_template.jinja in the repository. · Documentation and upkeep, The model card earns 11, missing limitations, languages and a knowledge cut-off.
- Sources: 24, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/minimax-m3-1-flash.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/minimax-m3-1-flash.svg` or a link to https://www.anchorterminal.com/tools/minimax-m3-1-flash from a page on minimax.io or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Check the use case first; commercial products must show "Built with MiniMax M3.1-Flash" and notify MiniMax once, or get written authorisation above $20M yearly revenue
2. On 8 H100 or H200 GPUs use vLLM's nightly image vllm/vllm-openai:minimax-m31-flash with --tensor-parallel-size 8 --block-size 128 --kv-cache-dtype fp8 --moe-backend marlin --max-model-len 262144
3. Pass --reasoning-parser minimax_m3 --tool-call-parser minimax_m3 --enable-auto-tool-choice in vLLM, or minimax-m3 for both in SGLang with --trust-remote-code
4. --block-size 128 is mandatory in vLLM, since MSA's sparse block size is 128
5. Set effort per request with chat_template_kwargs reasoning_effort (low, medium, high, xhigh, max); sample at temperature 1.0 and top_p 0.95
6. On Blackwell or MI355X, 4 GPUs serve it with the DSpark draft from dspark/; vLLM's Hopper command leaves DSpark off and turns video input off
7. For GGUF builds by others, use llama.cpp b11861 or later; earlier builds mis-scale the NVFP4 experts and produce garbage
8. Without the hardware, call MiniMax-M3.1-Flash on MiniMax's API (minimax-api), whose thinking defaults to max effort

## Connect

```bash
# vLLM's recipe, 8 x H100 or H200 (nightly image; no pip release supports it on 24 October 2026)
docker run --gpus all --ipc=host -p 8000:8000 -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:minimax-m31-flash MiniMaxAI/MiniMax-M3.1-Flash \
  --tensor-parallel-size 8 --block-size 128 --disable-hybrid-kv-cache-manager \
  --kv-cache-dtype fp8 --attention-config '{"indexer_kv_dtype": "fp8"}' --moe-backend marlin \
  --max-model-len 262144 --reasoning-parser minimax_m3 \
  --enable-auto-tool-choice --tool-call-parser minimax_m3
```

```bash
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "MiniMaxAI/MiniMax-M3.1-Flash", "temperature": 1.0, "top_p": 0.95, "chat_template_kwargs": {"reasoning_effort": "medium"}, "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ No safety results, no reporting route, remote code on one route (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ One Init commit and runtimes still in pre-release (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ The template documents the tool format and the card documents none of it (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Roughly 300 GB of weights and a licence with conditions (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)

- Arbiter's ruling (2026-10-25; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 3. Three reviewers note that only llama.cpp b11861 ran it as a release on 24 October 2026, with vLLM and SGLang on pre-release builds, and two note the licence's military ban and $20M threshold. Read it as a 250 GB multimodal model with a documented tool format in its template, no safety results or disclosure route, and runtime support still arriving.