Remote OpenClaw Blog
MiniMax M2 on OpenClaw: Setup, Pricing, and Performance Guide
9 min read ·
Remote OpenClaw Blog
9 min read ·
MiniMax is a Chinese AI company that has been building large language models since 2021, primarily focused on achieving high performance at low inference cost. Their M2 family is the latest product of that philosophy — a Mixture of Experts architecture that activates only 10 billion parameters per forward pass out of 230 billion total, resulting in a model that punches far above its compute weight class.
The pitch is simple: 90% of frontier model quality at 7% of the cost. That sounds like marketing, but the benchmarks largely back it up. M2.7, the latest version, scores 78% on SWE-bench Verified — within one percentage point of Claude Sonnet 4 — while costing $0.30 per million input tokens compared to Claude's $3.00.
For OpenClaw operators, MiniMax is especially interesting because the M2.5 variant is completely free on OpenRouter. That means you can test and run OpenClaw agent workflows without spending a single dollar on inference, then upgrade to M2.7 when you need the extra capability.
MiniMax has released three versions in the M2 family, each building on the same base architecture with incremental improvements:
| Version | Release | Key Improvement | Availability |
|---|---|---|---|
| M2 | Late 2025 | Original release: 230B MoE, 10B active, strong multilingual | HuggingFace, self-host |
| M2.5 | January 2026 | Improved instruction following, tool use, and safety alignment | OpenRouter (FREE), HuggingFace |
| M2.7 | March 2026 | Highest benchmarks (78% SWE-bench), improved reasoning chains | OpenRouter ($0.30/$1.20) |
The naming convention is straightforward: M2 is the base, and the decimal indicates the post-training iteration. M2.5 added better tool calling and instruction adherence. M2.7 pushed benchmark scores higher with improved reasoning and chain-of-thought capabilities.
For most OpenClaw operators, the decision is between M2.5 (free, good enough for testing and light workloads) and M2.7 (paid, best performance for production use).
| Specification | Value |
|---|---|
| Total Parameters | 230 billion |
| Active Parameters | ~10 billion per forward pass |
| Architecture | Mixture of Experts (MoE) |
| Inference Speed | ~100 tokens/second |
| Context Window | 128K tokens |
| Modalities | Text (M2.7 adds limited vision) |
| Tool Use | Yes (OAuth plugin support) |
| Developer | MiniMax (Beijing) |
The 10B active parameter count is the key to MiniMax's cost efficiency. While the model has access to 230B parameters worth of learned knowledge through its expert routing system, each inference pass only computes through 10B — roughly the size of a small Llama model. This is why MiniMax can price M2.7 at $0.30 per million input tokens while maintaining strong performance: the compute cost per token is genuinely low.
The ~100 tokens per second throughput is also notable. This is faster than many larger models (GLM-5 runs at ~69 tok/s, for comparison) and fast enough that users do not experience noticeable delays during real-time agent interactions.
Here are M2.7's headline benchmark results and how they compare to the models OpenClaw operators typically use:
| Benchmark | M2.7 | Claude Sonnet 4 | GPT-4.1 |
|---|---|---|---|
| SWE-bench Verified | 78% | ~79% | ~78% |
| HumanEval | 89.5% | ~92% | ~90% |
| MMLU | 86.8% | ~90% | ~89% |
| MATH | 84.2% | ~86% | ~85% |
The pattern is consistent: M2.7 lands within 1-3 percentage points of Claude Sonnet and GPT-4.1 across every major benchmark. It is not quite as capable on the margins — the gap shows up most on complex, multi-step reasoning chains and creative tasks — but for the vast majority of practical agent workflows, the difference is imperceptible.
The 78% SWE-bench Verified score is the most relevant metric for OpenClaw operators running coding agents. This means M2.7 can autonomously resolve roughly four out of five real-world GitHub issues — reading the codebase, understanding the bug, and generating a working fix. That is production-grade reliability for most software engineering tasks.
MiniMax's pricing is the strongest argument for using it with OpenClaw. Here is the full breakdown:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Monthly Cost (10M tokens/day) |
|---|---|---|---|
| MiniMax M2.5 | FREE | FREE | $0 |
| MiniMax M2.7 | $0.30 | $1.20 | ~$225 |
| Claude Sonnet 4 | $3.00 | $15.00 | ~$2,700 |
| GPT-4.1 | $2.00 | $8.00 | ~$1,500 |
The "90% quality at 7% cost" claim comes from dividing M2.7's OpenRouter cost by Claude Sonnet's: $0.30 / $3.00 = 10% input cost, $1.20 / $15.00 = 8% output cost. Blended, you are paying roughly 7-10% of what you would spend on Claude for an agent that scores within 1% on SWE-bench.
For an OpenClaw operator running a busy coding agent — processing 10 million tokens per day across dozens of tasks — the difference between $225/month (M2.7) and $2,700/month (Claude Sonnet) is substantial. That is over $2,400/month in savings, or nearly $30,000/year.
And if you are just getting started or running a low-volume agent, M2.5 at zero cost means you can operate indefinitely without any inference spending.
OpenRouter is the simplest way to connect MiniMax to OpenClaw. Both M2.5 (free) and M2.7 (paid) are available through the same API.
Sign up at openrouter.ai and generate an API key. For M2.5, you do not need to add credits — it is free. For M2.7, add at least $5 to start.
# In your OpenClaw config (e.g., ~/.openclaw/config.yaml)
llm:
provider: openrouter
model: minimax/m2.5-free
api_key: your-openrouter-api-key
temperature: 0.7
max_tokens: 8192# In your OpenClaw config (e.g., ~/.openclaw/config.yaml)
llm:
provider: openrouter
model: minimax/m2.7
api_key: your-openrouter-api-key
temperature: 0.7
max_tokens: 8192openclaw startA practical workflow for new operators: start with M2.5 to build and test your agent workflows at zero cost. Once you have validated that the agent is working correctly, switch to M2.7 for better reasoning and benchmark performance. The configuration change is a single line.
Marketplace
Free skills and AI personas for OpenClaw — browse the marketplace.
Browse the Marketplace →MiniMax M2 models support OAuth-based plugin authentication, which enables OpenClaw to use external tools and APIs through the model's native tool-calling capabilities. This is useful for workflows that require web browsing, database queries, or third-party API calls.
In your OpenClaw plugin configuration, define the OAuth credentials for the service you want to connect:
# In ~/.openclaw/plugins/my-tool.yaml
name: my-external-tool
auth:
type: oauth2
client_id: your-client-id
client_secret: your-client-secret
token_url: https://api.example.com/oauth/token
scopes:
- read
- write# In your OpenClaw config
plugins:
enabled:
- my-external-tool
auto_authenticate: true# Start OpenClaw and test a tool-calling workflow
openclaw start
# Then ask your agent to use the external tool
# MiniMax will automatically authenticate via OAuthMiniMax's tool-calling reliability is strong — M2.7 correctly formats and executes tool calls in the vast majority of cases. For complex multi-tool workflows where the agent needs to chain several API calls together, M2.7 handles sequencing and error recovery well.
MiniMax M2 is the right choice for your OpenClaw agent in these specific scenarios:
For a broader comparison of free model options, see Free API Models for OpenClaw.
MiniMax M2 has real limitations that you should evaluate before making it your primary agent backend:
For strategies to minimize API costs across all models, see Cheapest Way to Run OpenClaw.
Yes. MiniMax M2.5 is available at zero cost on OpenRouter with no API key billing. It is rate-limited and slightly less capable than M2.7, but for testing workflows or low-volume agent tasks, it works well. You can start with M2.5 for free and upgrade to M2.7 when you need the extra performance.
M2.7 scores 78% on SWE-bench Verified — roughly on par with Claude Sonnet 4. Where it differs is cost: M2.7 runs at $0.30 per million input tokens on OpenRouter versus $3.00 for Claude Sonnet. That means you get approximately 90% of the quality at about 7% of the cost. For coding and structured tasks, M2.7 is a strong alternative. For nuanced reasoning and creative writing, Claude still leads.
Yes. MiniMax supports OAuth-based plugin authentication, which means you can connect it to OpenClaw's plugin system for tool use, web browsing, and external API calls. Configure the OAuth credentials in your OpenClaw plugin settings and MiniMax will authenticate automatically when invoking tools.
M2 was the original release — a 230B total parameter MoE model with 10B active. M2.5 was a post-training improvement with better instruction following and tool use, offered free on OpenRouter. M2.7 is the latest version with the highest benchmark scores (78% SWE-bench), improved reasoning, and paid pricing at $0.30/$1.20 per million tokens. Each version builds on the same base architecture with incremental improvements.