Back to Models

MiniMax M2.1

minimaxminimax/m2.1

MiniMax M2.1 is an open-weight model focused on coding, tool use, and long-horizon task planning. Trained with emphasis on practical benchmarks covering front-end, backend, and workflow automation. General-purpose backbone for agent-based applications with reliable instruction following.

MiniMax M2.1 is an open-weight Mixture-of-Experts (MoE) language model released on December 23, 2025, as an incremental upgrade to the M2 family. With 230 billion total parameters but only 10 billion activated per token, it delivers a strong compute-to-knowledge ratio β€” rivaling closed-source coding models while remaining self-hostable on consumer-grade multi-GPU setups. MiniMax's stated focus for M2.1 was improving real-world complex task performance, especially in multi-language programming (Rust, Java, Golang, C++, Kotlin, TypeScript, etc.), full-stack web and mobile app development, office automation with composite instruction constraints, and agent/tool framework generalization.

For a TheRouter operator, M2.1 is the go-to open-weight model for agentic coding and software engineering workflows where cost-per-agent-turn matters. It scores competitively against much larger or closed models on SWE-bench Verified (74.0%), Multi-SWE-bench (49.4%), and the novel VIBE benchmark (88.6% aggregate), while its tool-use performance on Toolathlon (43.5%) and BrowseComp (47.4%) demonstrates real browser-and-shell agent capability. The model ships under a Modified-MIT license, enabling commercial use and self-hosting fallback. Its successor (M2.5, February 2026) and M2.7 (April 2026) have since shipped; for brand-new agent projects, evaluate those first, but M2.1 remains the most battle-tested, best-documented release in the family.

Best for
  • β€’ Agentic coding workflows β€” M2.1 achieves 74.0% on SWE-bench Verified across multiple agent frameworks (Claude Code, Droid, mini-SWE-agent) demonstrating consistent framework-agnostic performance
  • β€’ Multi-language development β€” systematic support for Rust, Java, Golang, C++, Kotlin, TypeScript, JavaScript, Objective-C, with 72.5% on SWE-bench Multilingual and 49.4% on Multi-SWE-bench
  • β€’ Full-stack app generation (VIBE) β€” the model scored 91.5 in Web, 89.7 in Android, 88.0 in iOS, 87.1 in Simulation, and 86.7 in Backend on the VIBE benchmark, supporting end-to-end 'zero to one' application scaffolding with visual and interactive quality checking
  • β€’ High-throughput agent serving β€” concise response style and reduced token consumption vs M2, combined with the 10B-active MoE architecture, make it cost-efficient for continuous agent workflows
Reach for something else if
  • β€’ Native multimodal input β€” M2.1 is text-in / text-out only (no vision). For tasks requiring image, video, or audio input, route to doubao-seed-2.0-vision or gpt-image-1.5
  • β€’ Deep math reasoning (AIME25: 83.0%) β€” while solid, models like DeepSeek V3.2 (92.0%), GPT-5.2 thinking (98.0%), or Gemini 3 Pro (96.0%) are significantly stronger for competition-level math
  • β€’ New agent project scaffolding β€” M2.5 (Feb 2026) and M2.7 (Apr 2026) have shipped as successors with further improvements; new integrations should start with M2.5 unless a known M2.1-specific regression makes it preferable
Context Length
205K
Max Output
66K
Input Priceper 1M tokens
$0.324/ 1M tokens
Output Priceper 1M tokens
$1.30/ 1M tokens

Modalities

text→text

Pricing Breakdown

TypeRate
Input$0.324 / 1M tokens
Output$1.30 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release date2025-12-23www.minimax.io β†—verified
Architecture230B total / 10B active β€” MoE (Mixture of Experts) with interleaved thinking; Transformer backbonegithub.com β†—verified
Training cutoffNot publicly disclosedunknown
LicenseModified-MIT (commercial use permitted)huggingface.co β†—verified
Supported inference backendsSGLang, vLLM, Transformers, MLX-LM, KTransformersgithub.com β†—verified
Successor (M2.5)MiniMax M2.5 (February 2026) β€” further improvements over M2.1; available on TheRouter as minimax/m2.5www.minimax.io β†—to verify
Latest successor (M2.7)MiniMax M2.7 (April 2026) β€” latest in the M2 line; available on TheRouter as minimax/m2.7huggingface.co β†—to verify

Benchmarks

BenchmarkDistributionScoreSource
SWE-bench Verified
Averaged across Claude Code, Droid, and mini-SWE-agent scaffoldings; 4-run average
74.0%%github.com β†—
Multi-SWE-bench
49.4%%github.com β†—
SWE-bench Multilingual
72.5%%github.com β†—
Terminal-bench 2.0
47.9%%github.com β†—
VIBE (Average)
MiniMax-proprietary benchmark β€” full-stack app generation, Agent-as-a-Verifier paradigm; 3-run average
88.6github.com β†—
VIBE-Web
91.5github.com β†—
VIBE-Android
89.7github.com β†—
VIBE-iOS
88.0github.com β†—
VIBE-Simulation
87.1github.com β†—
VIBE-Backend
86.7github.com β†—
Toolathlon
43.5%%github.com β†—
BrowseComp
With context management: 62.0%
47.4%%github.com β†—
AIME25
83.0%%github.com β†—
MMLU-Pro
88.0%%github.com β†—
GPQA-Diamond
83.0%%github.com β†—
SWE-bench Verified (Droid)
71.3%%github.com β†—
SWE-bench Verified (mini-swe-agent)
67.0%%github.com β†—

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "minimax/m2.1",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

Chat completion

M2.1 uses the standard chat completions endpoint. The model includes interleaved thinking β€” chain-of-thought reasoning tokens are interleaved with the response. These thinking tokens are not charged separately and are exposed inline in the response.

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/m2.1",
    "messages": [{"role": "user", "content": "Write a Rust function that deduplicates a vector"}]
  }'

More from minimax

Similar models

Cross-provider sibling models
deepseek/deepseek-v3.2
671B MoE (37B active) β€” strong open-weight rival in coding and agent tasks
qwen/qwen3-235b
Alibaba Qwen3 235B β€” competitive open-weight coding model in similar size class
doubao/doubao-seed-1.6

Not in the current catalog; the description below is historical.

ByteDance SEED 1.6 β€” strong agent and coding performance from another Chinese lab

Browse the current model catalog
cohere/command-a

Not in the current catalog; the description below is historical.

Cohere Command A β€” enterprise-focused open-weight model with strong tool use

Browse the current model catalog

News & changes

2025-12-23

MiniMax M2.1 released β€” multi-language coding focus

MiniMax released M2.1 as an open-weight MoE model with 230B total / 10B active parameters, targeting multi-language programming, full-stack app generation (VIBE benchmark), compound instruction following for office scenarios, and framework-generalizing agent tooling. The release included a novel VIBE benchmark using an Agent-as-a-Verifier paradigm for app quality assessment. Published alongside deployment guides for SGLang, vLLM, and Transformers.

re-authored by TheRouterminimax.io/news β†—

Frequently asked

Is MiniMax M2.1 free to use for commercial projects?

Yes. M2.1 is released under a Modified-MIT license which permits commercial use, modification, and distribution. Both the code in the GitHub repository and the model weights on Hugging Face carry this permissive license.

What hardware do I need to run M2.1 locally?

As a 230B-parameter model with 10B active parameters, M2.1 can run on multi-GPU setups. MiniMax recommends SGLang or vLLM for serving. With FP8 quantization, the model fits on 2-4 consumer GPUs (e.g., RTX 5090). MLX-LM is available for Apple Silicon Macs with unified memory (128 GB+ recommended). See the GitHub repo for detailed deployment guides.

Does M2.1 support Interleaved Thinking (chain-of-thought reasoning)?

Yes. M2.1 ships with interleaved thinking as a core feature. Reasoning tokens are interleaved inline with the final response (not hidden or output separately). This means the model's chain-of-thought is visible in the same message stream, and there is no separate 'thinking' parameter or additional cost for reasoning tokens beyond standard completion pricing.

How does M2.1 compare to Claude Sonnet 4.5 on coding?

MiniMax's published benchmarks show M2.1 matches or exceeds Claude Sonnet 4.5 on multilingual coding (72.5% vs 68% on SWE-bench Multilingual), Toolathlon (43.5% vs 38.9%), and VIBE aggregate (88.6 vs 85.2). Sonnet 4.5 leads on SWE-bench Verified (77.2% vs 74.0%). M2.1 is significantly cheaper β€” at $0.40/M input tokens on TheRouter vs Sonnet 4.5's pricing β€” making it a strong price-performance pick for agentic coding.

Should I use M2.1 or the latest M2.7 for new projects?

For brand-new projects, start with M2.7 (April 2026, available as minimax/m2.7 on TheRouter), which is the latest in the M2 family with self-evolution and advanced agent capabilities. Use M2.1 if you need the most battle-tested version with the widest deployment documentation, or if you have workflows that were tuned for M2/M2.1 and you want to avoid any regression risk from the later releases.

Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release datewww.minimax.io β†—2026-05-27verified
Architecturegithub.com β†—2026-05-27verified
Training cutoffβ€”β€”unknown
Licensehuggingface.co β†—2026-05-27verified
Supported inference backendsgithub.com β†—2026-05-27verified
Successor (M2.5)www.minimax.io β†—2026-05-27to verify
Latest successor (M2.7)huggingface.co β†—2026-05-27to verify
SWE-bench Verifiedgithub.com β†—2026-05-27verified
Multi-SWE-benchgithub.com β†—2026-05-27verified
SWE-bench Multilingualgithub.com β†—2026-05-27verified
Terminal-bench 2.0github.com β†—2026-05-27verified
VIBE (Average)github.com β†—2026-05-27verified
VIBE-Webgithub.com β†—2026-05-27verified
VIBE-Androidgithub.com β†—2026-05-27verified
VIBE-iOSgithub.com β†—2026-05-27verified
VIBE-Simulationgithub.com β†—2026-05-27verified
VIBE-Backendgithub.com β†—2026-05-27verified
Toolathlongithub.com β†—2026-05-27verified
BrowseCompgithub.com β†—2026-05-27verified
AIME25github.com β†—2026-05-27verified
MMLU-Progithub.com β†—2026-05-27verified
GPQA-Diamondgithub.com β†—2026-05-27verified
SWE-bench Verified (Droid)github.com β†—2026-05-27verified
SWE-bench Verified (mini-swe-agent)github.com β†—2026-05-27verified
MiniMax M2.1 released β€” multi-language coding focusminimax.io/news β†—2026-05-27verified
Is MiniMax M2.1 free to use for commercial projects?huggingface.co β†—2026-05-27to verify
What hardware do I need to run M2.1 locally?github.com β†—2026-05-27to verify
Does M2.1 support Interleaved Thinking (chain-of-thought reasoning)?www.minimax.io β†—2026-05-27to verify
How does M2.1 compare to Claude Sonnet 4.5 on coding?github.com β†—2026-05-27to verify
Help & contact