Overview
The March 18, 2026 MiniMax release introduces M2.7 — the first model in the M-series to participate in its own evolutionary development cycle. Where M2.5 focused on advancing coding benchmarks and office task automation, M2.7 shifts the architecture toward autonomous self-improvement: the model can update its own memory, build dozens of complex skills in its agent harness, and run reinforcement learning experiments without human intervention.
M2.7 supports a 204,800-token context window and runs at approximately 60 TPS in standard configuration, with a MiniMax-M2.7-highspeed variant reaching approximately 100 TPS at higher throughput cost. On software engineering benchmarks, M2.7 reaches 56.22% on SWE-Pro — near Claude Opus 4.6 parity — while the AI agent platform at agent.minimax.io gives non-technical users direct access without API setup.
What's New
Recursive Self-Improvement
M2.7 is the first MiniMax model designed to participate in its own training loop. It can update its own memory store, construct new harness skills (each up to 2,000+ tokens), and manage reinforcement learning experiments autonomously. In internal MiniMax testing, 100+ self-improvement iterations produced a 30% performance improvement — a capability that has no direct precedent in the M-series history.
MiniMax reports a 97% skill adherence rate while working with over 40 complex skills, each exceeding 2,000 tokens, meaning the self-built skills stay stable and consistent rather than drifting during extended agentic sessions.
Native Multi-Agent Collaboration
M2.7 introduces built-in support for agent teams with defined role boundaries, adversarial reasoning between agents, and behavioral differentiation. Agents in a M2.7-powered team can take on distinct personas, challenge each other's reasoning, and coordinate toward complex goals — a significant step beyond M2.5's single-agent optimization focus.
This native multi-agent design is reflected in benchmarks: M2.7 scores 62.7% on MM Claw (agent interaction benchmark), approaching Claude Sonnet 4.6 on that dimension.
Benchmark Performance
| Benchmark | M2.7 | Notes |
|---|---|---|
| SWE-Pro | 56.22% | Near Claude Opus 4.6 level |
| VIBE-Pro | 55.6% | Full project delivery |
| Terminal Bench 2 | 57.0% | Complex systems understanding |
| MM Claw | 62.7% | Agent interaction, near Sonnet 4.6 |
| GDPval-AA | 1495 ELO | Highest among open-source models |
| MLE Bench Lite | 66.6% avg | ML competition tasks |
Throughput Variants
M2.7 ships in two configurations:
- MiniMax-M2.7: ~60 TPS, 204,800-token context window, $0.30/M input / $1.20/M output
- MiniMax-M2.7-highspeed: ~100 TPS, same context and performance, $0.60/M input / $2.40/M output
Prompt caching is available separately at $0.06/M read tokens and $0.375/M write tokens, applicable to both variants.
Professional Office Integration
M2.7 maintains strong office-task performance with significantly improved Excel, PowerPoint, and Word editing for multi-round revisions; separately, MiniMax reports a 97% skill adherence rate across 40 complex skills exceeding 2,000 tokens. It handles finance-specific tasks including revenue modeling, research report generation, and assumption design from company financials.
Availability & Access
M2.7 is available on MiniMax Agent (agent.minimax.io) for interactive use and on the MiniMax API platform (platform.minimax.io) for programmatic access. Both the standard M2.7 and M2.7-highspeed variants are accessible via the API. A Coding Plan subscription is available at platform.minimax.io/subscribe/coding-plan for development teams needing high daily usage.
| Access Path | Details |
|---|---|
| MiniMax Agent | Free tier with credits; no API key required |
| API Pay-As-You-Go | $0.30/M input, $1.20/M output (M2.7 standard) |
| API Highspeed | $0.60/M input, $2.40/M output (M2.7-highspeed) |
| Prompt Caching | $0.06/M read, $0.375/M write (both variants) |
Pricing & Plans
MiniMax M2.7 uses pay-as-you-go API pricing with no required subscription tier for standard access.
| Plan | Model | Input | Output |
|---|---|---|---|
| API Standard | M2.7 | $0.30/M | $1.20/M |
| API Highspeed | M2.7-highspeed | $0.60/M | $2.40/M |
| Prompt Cache Read | Both | $0.06/M | — |
| Prompt Cache Write | Both | $0.375/M | — |
For teams needing higher usage volumes, the Coding Plan subscription at platform.minimax.io/subscribe/coding-plan provides a structured alternative to pure pay-as-you-go. Enterprise pricing and SLA terms are available by contacting MiniMax directly.
Best For
- Development teams building AI agent pipelines that require multi-agent orchestration and autonomous task execution at low API cost
- ML researchers using GLM-5 or similar open-weight models who want a frontier alternative with self-improvement capabilities
- Software engineering teams comparing M2.7 against GPT-5.3-Codex or Claude Opus 4.6 on SWE-bench class tasks
- Finance, research, and office automation workflows requiring multi-round document generation with high skill adherence
- Organizations evaluating open-source-friendly vendors given MiniMax's track record of releasing weights for prior M-series versions
FAQ
How is M2.7 different from M2.5?
M2.5 (February 2026) optimized for coding SOTA benchmarks (80.2% SWE-Bench Verified) and faster agentic task completion. M2.7 shifts the design toward autonomous agent teams and recursive self-improvement — M2.7's 56.22% SWE-Pro result and M2.5's 80.2% SWE-Bench Verified result come from different benchmarks and should not be compared directly, but M2.7 adds native multi-agent collaboration, a 30% self-improvement gain through recursive iterations, and significantly more complex skill management capabilities.
What does "recursive self-improvement" mean in practice?
M2.7 can build its own harness skills (reusable instruction sets up to 2,000+ tokens each), update its own memory store based on task outcomes, and run RL experiments to optimize its own performance. In MiniMax's internal trials, 100+ automated iterations produced a 30% performance improvement. This makes M2.7 useful for tasks where you want the model to improve its approach over time rather than starting fresh each session.
Is M2.7 available as open-source weights?
MiniMax has released open weights for M1, M2, M2.1, and M2.5 on Hugging Face. Check the MiniMaxAI organization on Hugging Face for the current status of M2.7 weight availability — this is not yet confirmed as publicly released as of March 2026.
How does M2.7 pricing compare to Claude Opus 4.6?
MiniMax-M2.7 is priced at $0.30/M input and $1.20/M output. Claude Opus 4.6 is priced at approximately $5.00/M input. MiniMax says M2.7's 56.22% SWE-Pro score nearly approaches Opus's best level, but exact cross-model parity should be framed cautiously given the comparison relies on MiniMax-cited evaluations. For teams where API cost is a primary constraint and SWE-class performance is the target, M2.7 offers a substantially lower cost at comparable benchmark performance.
Can I use M2.7 without API access?
Yes. MiniMax Agent at agent.minimax.io provides free-tier access to M2.7 for interactive use — no API key or programming required. This makes it accessible for product designers, researchers, and users who want to test M2.7's agentic capabilities without setting up API billing.
