MiniMax M2.7

Supported Functionality

Item	Specification
Input	Text
Output	Text
Context	204,800 tokens (~205K)
Max Output	131,072 tokens (~131K)
Vision	✗ Not Supported (extendable via external tools / MCP)
Function Calling	✓ Supported (Agent Teams, dynamic tool search, MCP)

Description

MiniMax M2.7 is MiniMax's self-evolving large language model released on March 18, 2026 — a comprehensive upgrade over M2.5. It is a Mixture-of-Experts (MoE) model with 230 billion total parameters and just 10 billion activated per token, using top-k routing across 256 experts plus RoPE positional embeddings and QK RMSNorm for stable training. API pricing is $0.30 / $1.20 per million input/output tokens, with two variants: standard M2.7 and M2.7-highspeed. Weights are published on Hugging Face (MiniMaxAI/MiniMax-M2.7) under a non-commercial license — commercial use requires a separate agreement.

M2.7's defining breakthrough is being the first model that deeply participates in its own evolution. An internal M2.7 instance autonomously optimized a programming scaffold over 100+ rounds — analyzing failure trajectories, modifying code, running evaluations, and deciding to keep or revert — yielding a 30% performance gain. The model also built MiniMax's research agent harness with data pipelines, training environments, cross-team collaboration, and persistent memory, handling 30–50% of the RL research workflow. Benchmark-wise, M2.7 scores 56.22% on SWE-Pro (near Opus's best), 55.6% on VIBE-Pro, 57.0% on Terminal-Bench 2.0, 66.6% on MLE Bench Lite (trailing only Opus 4.6 and GPT-5.4), and an open-source-leading 1495 Elo on GDPval-AA — reducing live production incident recovery time to under three minutes.

Key Capabilities

• Self-Evolution & Recursive Improvement: First model to deeply participate in its own post-training; auto-improved a programming scaffold across 100+ rounds for a 30% capability gain.
• Native Agent Teams: Multi-agent coordination is internalized — no external orchestration code needed for planning/execution/validation/correction loops, with stable role identity.
• Dynamic Tool Search: Discovers and invokes tools at runtime based on evolving task state rather than relying on declared manifests, expanding the effective action space.
• Real-World Software Engineering: SWE-Pro 56.22%, SWE Multilingual 76.5, Multi-SWE Bench 52.7, VIBE-Pro 55.6%, NL2Repo 39.8% — strong end-to-end project delivery.
• SRE & Systems-Level Reasoning: Correlates monitoring metrics, trace analysis, and database root-cause checks; consistently cuts live incident recovery to under 3 minutes.
• Top Professional Office Skills: GDPval-AA Elo 1495 leads open-source; significantly stronger multi-turn high-fidelity editing across Excel, PPT, and Word than M2.5.
• Complex Skill Adherence: Maintains a 97% skill-following rate across 40 cases of complex Skills exceeding 2000 tokens — exceptional long-instruction stability.

Technical Strengths

Feature	Benefit
230B MoE / 10B Active	Large-scale knowledge with inference cost close to a 10B dense model; $0.30/$1.20 per million tokens is highly competitive
Top-k Expert Routing + QK RMSNorm	Stable training at scale; only relevant experts activate per token, keeping per-token compute manageable
Native Agent Teams Orchestration	Multi-agent coordination is baked into the model, eliminating the need for external harness scaffolding
Dynamic Tool Search	Tools no longer bound by upfront manifests — the model discovers and calls new tools as tasks evolve
M2.7 + M2.7-highspeed Variants	Standard and ~100 t/s high-throughput versions produce identical quality — choose by latency/cost needs
Open Weights + Deep NVIDIA Optimization	Weights on HF/ModelScope; NVIDIA collaborates with vLLM/SGLang for 2.5–2.7× throughput gains and FP8 MoE kernels

Capability Ratings

Dimension	Rating	Notes
Reasoning	Excellent	AA Intelligence Index 50; MLE Bench Lite 66.6% trails only Opus 4.6 and GPT-5.4
Coding	Top-tier	SWE-Pro 56.22 matches GPT-5.3-Codex; SWE Multilingual 76.5 and Terminal-Bench 2.0 57.0
Creative Writing	Strong	Solid character consistency and emotional intelligence, though tuning skews toward agents and engineering
Multimodal	Moderate	Text-only natively; vision/audio require external tools or MCP
Response Speed	Moderate	Standard ~48 t/s (verbose outputs); M2.7-highspeed ~100 t/s; full attention slows long-context workloads
Context Window	Large	204,800 tokens with 131K max output — covers most codebases and long-document scenarios

Use Cases

• AI Coding Assistants & End-to-End Engineering: SWE-Pro 56.22 plus native Agent Teams make it a cost-effective backend for OpenCode, Cline, Cursor, and similar tooling.
• SRE & Production Incident Response: Autonomously correlates monitoring, logs, and code repositories, consistently cutting incident recovery to under 3 minutes.
• Automated ML Research: 66.6% on MLE Bench Lite — handles 30–50% of research pipelines (literature review, data pipelines, training, debugging, merges).
• Professional Office Automation: #1 open-source on GDPval-AA (1495 Elo) with high-fidelity multi-turn Excel/PPT/Word editing for finance, reporting, and knowledge work.
• Complex Multi-Agent Workflows: Native Agent Teams enable plan/execute/validate/correct loops without external orchestration code.
• Dynamic Tool-Discovery Agents: Autonomous discovery and invocation of undeclared tools fits enterprise environments where tool catalogs shift rapidly.
• Self-Hosted Enterprise Deployment: Open weights plus NVIDIA NIM endpoints and vLLM/SGLang support enable on-prem inference (commercial use requires a separate license).