MiniMax M2.7 is a 230-billion-parameter MoE model that activates only 10 billion parameters per token. It scores 56.22% on SWE-Pro, runs at 100 tokens/second, and costs $0.30 per million input tokens. That’s 50x cheaper than Claude Opus.

Architecture

Spec	Value
Total parameters	230B
Active parameters	10B per token (4.3% activation)
Architecture	Sparse MoE (256 experts, 8 active per token)
Layers	62
Context window	200K tokens
Speed	~100 tokens/second
Attention	Multi-head causal with RoPE + QK RMSNorm
License	Open weights

The extreme sparsity (4.3% activation) is what makes M2.7 so cheap to run. Only 10B of 230B parameters compute for each token, but the model has access to the full 230B knowledge base through expert routing.

Benchmarks

Benchmark	M2.7	Claude Opus 4.6	GPT-5.4	DeepSeek V3
SWE-Pro	56.22%	57.3%	57.7%	~54%
SWE-bench Verified	~78%	80.8%	—	77.8%
Coding (Kilo Code test)	~90% of Opus	Baseline	—	~85% of Opus

M2.7 is within 1-2 points of Claude Opus on SWE-Pro. In real-world coding tests, it delivers roughly 90% of Opus quality.

Pricing

	MiniMax M2.7	Claude Opus	DeepSeek	GLM-5.1
Input	$0.30/1M	$15.00/1M	$0.27/1M	$1.00/1M
Output	$1.20/1M	$75.00/1M	$1.10/1M	$2.30/1M
Cost for 1hr coding	~$0.50	~$25	~$0.45	~$2

M2.7 and DeepSeek are the two cheapest frontier-class models. M2.7 is slightly more expensive but faster (100 tok/s vs 60 tok/s).

The self-evolving feature

M2.7’s headline feature is “self-evolution” — it uses multi-agent collaboration internally to plan, execute, and refine tasks. In practice, this means it can:

• Break complex tasks into subtasks
• Execute them with different internal strategies
• Evaluate results and retry with different approaches
• Improve its output iteratively within a single session

This is similar to Kimi K2.5’s Agent Swarm but happens internally rather than through external parallel agents.

How to use M2.7

Via OpenRouter

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your-openrouter-key"
)

response = client.chat.completions.create(
    model="minimax/minimax-m2.7",
    messages=[{"role": "user", "content": "Refactor this function to use async/await"}]
)

See our OpenRouter guide for full setup.

With Aider

aider --model openrouter/minimax/minimax-m2.7

See our Aider guide.

With OpenCode

opencode --model minimax/minimax-m2.7

See our OpenCode guide.

Via MiniMax API directly

client = OpenAI(
    base_url="https://api.minimax.chat/v1",
    api_key="your-minimax-key"
)

response = client.chat.completions.create(
    model="MiniMax-M2.7",
    messages=[{"role": "user", "content": "Write a REST API in Express"}]
)

When to use M2.7

Best for:

• Budget-conscious teams needing near-frontier quality
• High-volume coding tasks (the speed + low cost combo)
• Agentic workflows (self-evolving capability)
• Replacing expensive Claude Opus calls for routine work

Not ideal for:

• Tasks requiring absolute best quality (Claude Opus still leads by ~10%)
• Self-hosting (no open weights for local deployment yet)
• GDPR-sensitive workloads (Shanghai-based servers)

M2.7 vs M2.5

	M2.7	M2.5
Released	March 2026	February 2026
SWE-bench	~78%	80.2%
SWE-Pro	56.22%	—
Self-evolving	✅	❌
Speed	100 tok/s	Slower
Price	$0.30/1M	$0.15/1M

M2.5 actually scores higher on SWE-bench Verified but M2.7 has the self-evolving capability and is faster. For pure coding, M2.5 might be the better value at half the price.

Bottom line

MiniMax M2.7 is the best value proposition in AI coding right now. 90% of Claude Opus quality at 2% of the cost. If you’re using model routing (cheap model for routine work, expensive model for hard problems), M2.7 should be your cheap model.

Related: What is MiniMax? · MiniMax M2.7 vs Claude vs DeepSeek · How to Use MiniMax API · How to Reduce LLM API Costs