MiniMax Unleashes M2.7: The Open-Source Model That Rewrites Its Own Code
News Summary
On April 12, 2026 (CST), AI company MiniMax officially open-sourced its flagship agent model MiniMax M2.7, making the full model weights publicly available on Hugging Face. M2.7 is MiniMax's most capable open-source release to date and the first model in the series to actively participate in its own training and optimization cycle — a capability the company describes as "self-evolution." The model scored 56.22% on the SWE-Pro benchmark and 57.0% on Terminal Bench 2, placing it among the top-performing open-source models in software engineering tasks.
Model Architecture and Technical Specifications
MiniMax M2.7 is built on a sparse Mixture-of-Experts (MoE) architecture, featuring 230 billion total parameters with only 10 billion activated per token. It supports a context window of 200,000 tokens and includes 256 expert modules. The model uses multi-head causal self-attention enhanced with Rotary Position Embeddings (RoPE) and Query-Key Root Mean Square Normalization (QK RMSNorm) for stable training at scale. A top-k expert routing mechanism ensures that only the most relevant experts are engaged per input, keeping inference costs low despite the model's large overall parameter count.
Self-Evolution: A New Training Paradigm
One of M2.7's most distinctive features is its capacity to participate in its own development. Built on top of MiniMax's internal agent workflow framework OpenClaw, the model executes an iterative self-improvement loop: after completing a task batch, it analyzes failure trajectories, identifies specific error points, modifies its own scaffold code, re-runs evaluations, and then decides whether to retain or roll back changes. In one documented run targeting programming performance optimization, M2.7 ran this loop autonomously for over 100 rounds. The process yielded a 30% improvement on internal evaluation sets without human intervention. Among the strategies M2.7 independently discovered were systematic sampling parameter tuning — optimizing temperature, frequency penalty, and presence penalty — and the automatic propagation of bug fixes across similar code patterns in other files.
Software Engineering and Professional Office Performance
In real-world software engineering contexts, M2.7 demonstrated strong capabilities across end-to-end project delivery, log analysis, bug diagnosis, code security review, and machine learning workflow debugging. On the SWE-Pro benchmark, it achieved a score of 56.22%, matching GPT-5.3-Codex. On VIBE-Pro, which evaluates complete repository-level code generation across Web, Android, and iOS targets, it scored 55.6%. For complex system comprehension, Terminal Bench 2 yielded a score of 57.0% and NL2Repo returned 39.8%.
In professional office scenarios, M2.7 ranked first among open-source models on the GDPval-AA evaluation with an ELO score of 1,495. The model supports sophisticated multi-round editing of Excel, PowerPoint, and Word documents, and can autonomously build revenue forecasting models from annual report data, generate Excel pivot tables, draft Word research reports, and produce presentation decks — a workflow the company describes as equivalent to junior analyst output.
Multi-Agent Coordination and Skill Execution
M2.7 natively supports multi-agent coordination without requiring external orchestration frameworks. Using Agent Teams, complex Skills, and a dynamic Tool Search mechanism, the model can decompose long-horizon tasks and distribute execution across coordinated agents. In tests involving 40 complex skills requiring over 2,000 tokens per case, M2.7 maintained a 97% skill-following rate. On MM-Claw, MiniMax's proprietary real-task evaluation set drawn from OpenClaw workflows, M2.7 scored 62.7%.
OpenRoom Interactive System
Alongside M2.7's open-source release, MiniMax also published an interactive system called OpenRoom, which extends AI interaction beyond text into a visual, web-based GUI environment. OpenRoom supports real-time scene feedback and dynamic human-AI interaction, and is designed to be highly extensible for future human-computer interaction modalities.
Hardware Ecosystem and Inference Deployment
On the day of open-source release (April 12, 2026, CST), MiniMax announced completed inference adaptation with a broad set of hardware and software partners. NVIDIA published deployment guidance for M2.7 on its technical blog (April 11, 2026, EDT), introducing NVIDIA NemoClaw — an open-source reference stack for running OpenClaw agents securely on NVIDIA Blackwell Ultra GPUs, with optimizations for both vLLM and SGLang serving frameworks delivering up to 2.7x throughput improvement. The model is also adapted for Huawei Ascend (Atlas 800 A3 and A2 series via vllm-Ascend), Moore Threads (MUSA architecture, MTT S5000), and other silicon partners. On the software side, platforms including Together AI, Fireworks, Ollama, vLLM, SGLang, and community ecosystems such as ModelScope have integrated M2.7 as of launch day.
Availability
MiniMax M2.7 open weights are available on Hugging Face at MiniMaxAI/MiniMax-M2.7. API access is available through the MiniMax developer platform at platform.minimax.io. The model is also accessible via the MiniMax Agent product and is already embedded in popular agent tools including Hermes Agent and OpenClaw.