What is MiniMax-M2.7?
MiniMax-M2.7 is a 228.7B-parameter open-weight model from MiniMax, published on Hugging Face on April 9, 2026. MiniMax describes it as the first of its models to take a hand in its own development: during training it updated its own memory, built dozens of complex skills for RL experiments, and changed its own learning process based on the results. One internal version rewrote a programming scaffold over 100+ rounds, analyzing failure trajectories, editing code, running evaluations, and deciding to keep or revert each change, and came out 30% better for it. The released model is built for the same kind of work: complex agent harnesses, long productivity tasks, native Agent Teams for multi-agent collaboration, and dynamic tool search.
| Specification | MiniMax-M2.7 |
|---|---|
| Total parameters | 228.7B |
| Modalities | Text in, text out |
| Agent features | Agent Teams, complex Skills, dynamic tool search |
| Recommended serving | SGLang, vLLM or Transformers; also on NVIDIA NIM |
| Recommended sampling | temperature 1.0, top_p 0.95, top_k 40 |
| Release date | April 9, 2026 |
| License | Custom, listed as "other" on Hugging Face |
The card positions M2.7 well beyond code generation. MiniMax claims system-level engineering skills: correlating monitoring metrics, running trace analysis, verifying root causes in databases, and making SRE-level decisions, and says the model has cut live production incident recovery to under three minutes on multiple occasions. On the office side it edits Word, Excel and PPT files over multiple rounds while keeping the deliverables editable, and MiniMax reports 97% skill compliance across 40+ complex skills on MM Claw.
MiniMax-M2.7 benchmarks
All numbers below are MiniMax's own, from the model card. The vendor published no side-by-side competitor table, so the scores stand alone: MLE Bench Lite is a medal rate across 22 ML competitions, GDPval-AA is an ELO rating, the rest are percentages.
| Benchmark | MiniMax-M2.7 |
|---|---|
SWE-Pro Harder engineering | 56.22 |
SWE Multilingual Multilingual engineering | 76.5 |
Terminal Bench 2 Terminal agents | 57.0 |
NL2Repo Repo-level coding | 39.8 |
VIBE-Pro Vibe coding | 55.6 |
Toolathon Long-horizon tools | 46.3 |
MLE Bench Lite ML competitions | 66.6 |
GDPval-AA Professional tasks | 1495 |
MiniMax's framing for these: SWE-Pro matches GPT-5.3-Codex, VIBE-Pro is nearly on par with Opus 4.6, and the GDPval-AA ELO is the highest among open-weight models, ahead of GPT5.3. The card also concedes ground: the MLE Bench Lite medal rate is second to Opus-4.6 and GPT-5.4, and on the MM Claw end-to-end benchmark its 62.7% lands close to, not above, Sonnet 4.6.
MiniMax-M2.7 hardware requirements
The system requirement to check is memory. The community GGUF builds come from unsloth/MiniMax-M2.7-GGUF, shipped as multi-part files, so the sizes below are per-build totals. The unquantized BF16 conversion is 457.5 GB; the quantized ladder is what makes local use realistic.
| Memory | Build to pick | File size |
|---|---|---|
| 64 GB | UD-IQ1_M | 60.7 GB |
| 96 GB | UD-IQ3_S | 83.6 GB |
| 128 GB | UD-IQ4_XS | 108.4 GB |
| 192 GB | UD-Q5_K_XL | 169.5 GB |
| 256 GB and up | Q8_0 | 243.1 GB |
When two builds both fit, take the larger one. That matters most in the 1-bit and 2-bit range, where quality falls fastest per gigabyte saved: on a 96 GB machine UD-IQ3_S at 83.6 GB is worth the squeeze over UD-IQ2_M at 70.2 GB. If quant names and split files are new territory, start with what GGUF is.
How to run MiniMax-M2.7 in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for MiniMax-M2.7 in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the rest of the lineup, see every MiniMax model you can run locally, or the sibling MiniMax M3.
MiniMax-M2.7 license
MiniMax-M2.7 ships under MiniMax's own license: Hugging Face lists it as "other" rather than a standard permissive license, and the full text lives in the LICENSE file of the model repository. Read that file before you build a commercial product on the model; the terms are the vendor's, not Apache or MIT.