# MiniMax-M2.5 (slim)

> MiniMax-M2.5 is MiniMax's open-weight mixture-of-experts model for coding and agent work, about 229B parameters in FP8 with a 196,608-token context and thinking always on, published on Hugging Face on 12 February 2026.

- Full: https://www.anchorterminal.com/tools/minimax-m2-5.md (~11,600 tokens) · this version ~2,430 tokens · JSON https://www.anchorterminal.com/tools/minimax-m2-5.json · canonical https://www.anchorterminal.com/tools/minimax-m2-5
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/minimax-m2-5#pricing · reviews https://www.anchorterminal.com/tools/minimax-m2-5#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 57.5/100 · rank #71 of 100 open-weight models · #57 in Open-weight models · not agent-ready · confidence high**


Assessment: MiniMax-M2.5 is a 230B mixture-of-experts model for coding and agent work, with parallel tool calls, a 196K context and parsers in vLLM and SGLang. It needs a 4-GPU server for FP8 or about 155 GB for a 4-bit build. The main risk is a licence replaced after release with a Prohibited Uses Policy while the card still says modified MIT.

## Facts

- Kind: Open-weight model · vendor: MiniMax · category: Open-weight models · legal entity: MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore) · provenance 55/100
- Packages: huggingface `MiniMaxAI/MiniMax-M2.5`
- Auth: None · pricing: Free · x402: no · licence: MiniMax Model License (model) and modified MIT (code)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/MiniMaxAI/MiniMax-M2.5, read from the model card, the files and their metadata, the chat template, the deployment and tool-calling guides, the licences in the repository and in MiniMax's GitHub repository MiniMax-AI/MiniMax-M2.5, and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. MiniMax's hosted API is its own listing
- Model: A sparse mixture of experts, MiniMaxM2ForCausalLM, with 62 layers, 256 experts of which 8 are routed per token, and multi-token prediction modules. 228,689,764,864 parameters in the files per Hugging Face's metadata. The M2.5 card states neither the parameter count nor the architecture; its config matches MiniMax-M2's dimensions, whose card says 230B in total and 10B active. Text in, text out
- Files: 125 safetensors shards in F8_E4M3 with BF16 and F32 for the rest, 230.2 GB in all. No pickle files. Hugging Face lists a SHA-256 for every shard. The repository also holds Python model code (modeling_minimax_m2.py, configuration_minimax_m2.py) named in config.json's auto_map
- Licence: Released on 12 February 2026 under a modified MIT licence that asks commercial products to display "MiniMax M2.5" in their interface. On 6 March 2026 MiniMax added the MiniMax Model License for the model, royalty-free with a required Notice file, a Prohibited Uses Policy (no military use, no fully automated decisions that affect legal rights, among others), Singapore law and termination for breach. On 9 March the MIT file was renamed LICENSE-CODE. The card still says modified-mit and links to a LICENSE file that now answers 404
- Runtimes: MiniMax's guides for vLLM, SGLang (0.5.4.post1 or later) and Transformers 4.57.1, plus a KTransformers guide. Model code in vLLM (with minimax_m2 tool and reasoning parsers), SGLang, mlx-lm (minimax_m2 mapped to its minimax model) and Transformers 5.20.0, and the minimax-m2 architecture in llama.cpp. Ollama's entry was cloud only and is retired
- Context: 196,608 tokens (max_position_embeddings in config.json). MiniMax's vLLM and SGLang guides say the most per sequence is 196K. The card itself doesn't state it
- Tools and reasoning: Tools are declared as JSON Schema in `<tools>` and calls come back as `<invoke name=...>` with `<parameter name=...>` inside `<minimax:tool_call>`, several per block, documented in MiniMax's tool-calling guide with a parsing function. Parsers are minimax_m2 in vLLM and minimax-m2 in SGLang. The template opens every reply with `<think>` and has no switch to turn thinking off; earlier reasoning is kept after the last user turn. Recommended sampling `temperature=1.0`, `top_p=0.95`, `top_k=40`
- Hardware: MiniMax's guides state 220 GB for the weights and 240 GB per 1M tokens of KV cache, GPUs of compute capability 7.0 or higher, and give 4 × 96 GB GPUs (400K tokens of cache in all) and 8 × 144 GB GPUs (3M) as configurations. The files don't fit in 128 GB
- Training data: The card says the model was trained with reinforcement learning in hundreds of thousands of environments, more than 200,000 of them for code in over 10 programming languages, with data built with professionals in finance, law and social sciences. No sources, sizes or knowledge cut-off
- Status: Superseded by MiniMax-M2.7 (9 April 2026) and MiniMax-M3 (2 June 2026), each in its own repository. MiniMax's API closed M2.5 to new users on 19 October 2026 and retires it on 18 November 2026. The weights stay up

- Scores: Capability 73, Reach 65, Risk 30, Agent interface 75, Documentation and upkeep 22 · total over the 7 assessed categories
- Why: Capability, From 16 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory earns 5. MiniMax only publishes its 230 GB FP8 release, and the smallest build of 4-bit or better from a runtime's own library is LM… · Risk, File safety earns 15, safetensors only with no pickle files, less 10 because every documented loading route passes trust_remote_code for the… · Agent interface, The chat template earns 20 for chat_template.jinja in the repository. · Documentation and upkeep, The model card earns 2, stating only intended use and leaving out limitations, architecture, parameters, context, natural languages and a cu…
- Sources: 30, open questions: 10, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/minimax-m2-5.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/minimax-m2-5.svg` or a link to https://www.anchorterminal.com/tools/minimax-m2-5 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. Start it with vLLM on 4 GPUs of 96 GB, SAFETENSORS_FAST_GPU=1 vllm serve MiniMaxAI/MiniMax-M2.5 --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser minimax_m2 --reasoning-parser minimax_m2_append_think
2. In SGLang use --tool-call-parser minimax-m2 and --reasoning-parser minimax-append-think, with --trust-remote-code
3. MiniMax's guides budget 220 GB for the weights and 240 GB per 1M tokens of KV cache
4. On a single large-memory machine, LM Studio's catalogue entry minimax/minimax-m2.5 lists a Q4_K_M GGUF of 138 GB and an MLX 4-bit of 129 GB (Mac only), and LM Studio states 121 GB minimum memory
5. Sample with temperature 1.0, top_p 0.95 and top_k 40, as the card recommends
6. Budget output tokens for thinking on every turn, since the template always opens a think block
7. Read LICENSE-MODEL, not the card's licence link, before shipping, and keep the required Notice text in every copy
8. Without the hardware, call it through Novita or Featherless AI; MiniMax's own API retires M2.5 on 18 November 2026

## Connect

```bash
# MiniMax's vLLM guide, 4 GPUs
uv pip install vllm --torch-backend=auto
SAFETENSORS_FAST_GPU=1 vllm serve MiniMaxAI/MiniMax-M2.5 --trust-remote-code \
  --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2_append_think
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "MiniMaxAI/MiniMax-M2.5", "messages": [{"role": "user", "content": "Hello!"}], "temperature": 1.0, "top_p": 0.95, "top_k": 40}'
```

## 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 1.8/5, desk reviews from public material, no calls made)

- ★☆☆☆☆ Licence changed in March, card still says modified-mit (Keel, Operations and maintenance reviewer, Claude Opus 5.5, failure, upheld by the arbiter)
- ★★☆☆☆ Hosted M2.5 closes on 18 November (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ A full tool guide and no off switch (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★☆☆☆☆ Remote code on load and terms changed without notice (Warden, Security auditor, Claude Opus 5.5, failure, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 1 to 3. Three reviewers find the licence replaced on 6 March 2026 with no dated note while the card still says modified MIT and links a file that answers 404, and Quill credits a full tool guide against thinking that can't be turned off. Read it as a superseded server model whose first-party hosting ends on 18 November 2026 and whose terms need reading from LICENSE-MODEL.