MiniMax M2.7

**MiniMax M2.7** is the April 2026 release from Chinese AI lab MiniMax. It is technically comparable to Claude Sonnet 4.6 on most benchmarks and within points of Opus 4.6 on several — but the announcement is overshadowed by a license change that has drawn substantial community pushback. ## Specifications - **230 billion parameter MoE** (Mixture of Experts) - **10 billion active parameters per token** - **256 local experts, 8 activated per token** - **62 layers, hidden size 3072** - **204,800 token context length** - **Not multimodal** (text only) - **Reasoning model** (chain-of-thought trained, uses reasoning tokens) - **SWE-Bench Pro**: 56.22% - **GDPval-AA**: top open-model ELO 1495 - Predecessors: MiniMax M2 (October 2025), M2.5 (February 2026), M2.7 (April 2026) ## Release timeline - **March 18, 2026**: M2.7 announced - **April 11-12, 2026**: weights dropped on HuggingFace ## NVIDIA partnership MiniMax ships a dedicated **NVIDIA NIM microservice** with FP8 MoE TensorRT-LLM optimizations, achieving **2.5-2.7x throughput improvements on Blackwell Ultra** GPUs versus unoptimized inference. This is a significant capability — for enterprise users with Blackwell infrastructure, M2.7 is one of the better-tuned open-weight models to deploy. ## License controversy **The big underreported story**: M2.7's license changed from M2/M2.5's permissive **MIT** to **commercial-use-restricted**. Written authorization from MiniMax is now required for commercial use. Non-commercial use (research, personal projects, fine-tuning for own use) remains free. **MiniMax's justification**: 'prevent service providers from nerfing the model when offering it to developers' — effectively, they don't want third-party hosts running quantized/limited versions and undercutting MiniMax's own commercial inference. **Community backlash was immediate and substantial**: - 'Open source my ass — they are liars' (top HuggingFace discussion). - Decrypt article: 'MiniMax Drops State-of-the-Art AI Agent Model — Then Quietly Changes the License.' - Per OSI/FSF definitions, this is no longer open source — it's **open-weight with commercial restrictions.** - MiniMax sometimes labels the license 'Modified-MIT,' which is misleading — MIT permits commercial use by definition. Calling it MIT while restricting commercial use is at best confusing, at worst deceptive. See Open Source vs Open Weight Debate for the broader naming distinction. ## Hardware reality - Full precision FP16: ~half a terabyte storage - **Q8 quantization**: ~261 GB - **Q4**: starts to lose accuracy - **Q2**: workable for a 230B model, accuracy hit acceptable - **Q1**: 'almost might as well not run it' Realistic local hardware for running M2.7: - Mac Studio / Mac Mini with 128GB+ unified memory - NVIDIA DGX Spark - AMD Strix Halo with 128GB unified memory - Server farm for full FP16 At release time no GGUF quantizations existed yet; Unsloth and community producers typically publish quantized versions within days for use with llama.cpp, AnythingLLM, LM Studio, Ollama. ## 'Self-evolving' marketing MiniMax's marketing describes M2.7 as 'our first model to deeply participate in its own evolution.' Practical translation: trained specifically to build the tools and scripts it needs to do a job. Same use-case space as Claude Code / OpenClaw and GLM 5.1 Open-Weight Model. The model writes small scripts to accomplish subtasks rather than solving everything in a single pass. ## MMClaw benchmark MiniMax's own evaluation benchmark, designed around tasks you'd do in OpenClaw. Per Tim Carambat's analysis, this is where gaps between M2.7 and Opus become visible — M2.7 performs at Sonnet-4.6-level on coding agent tasks. Pattern recognition: every major Chinese open-weight release in the past ~6 months (MiniMax, GLM, DeepSeek, Qwen) has positioned itself directly against Claude, not GPT. ## Context in the broader week April 7-12, 2026 window: - **April 7**: GLM 5.1 Open-Weight Model (ZAI, 754B MoE, MIT license) — beats GPT-5.4, Opus, Gemini on SWE-Pro. - **April 8**: Meta Muse Spark and the End of the Llama Era (closed) — Meta abandons open source. - **April 7-12**: Claude Mythos Preview (closed-partner-only) — cybersecurity capabilities, $100M Glasswing. - **April 11-12**: **MiniMax M2.7** (230B MoE, license restricted). See AI News Week of April 12 2026 — Four Headline Stories. Trend: Chinese labs keep shipping frontier-comparable open-weight models but are **starting to tighten licenses** (M2.7 is the clearest example). Commercial use of 'open' Chinese models is becoming murkier. ## Practical choice guidance For builders: GLM 5.1 Open-Weight Model (MIT, 754B) is more permissive than M2.7 (restricted, 230B). For end-users running locally: M2.7 may be more practical given hardware — 230B vs 754B = dramatically less RAM required. For research: both are fine non-commercial.