MiniMax M2.1

// Overview

MiniMax M2.1 is an open-weight large language model from MiniMax, released on December 23, 2025 as an enhanced revision of MiniMax M2. It is a sparse Mixture-of-Experts (MoE) model with roughly 230 billion total parameters but only about 10 billion activated per token, which keeps inference fast and cheap while retaining the capacity of a much larger dense model. The weights ship on Hugging Face under a modified MIT license, so teams can run M2.1 locally with SGLang, vLLM, Transformers, or KTransformers instead of relying solely on the hosted API.

The headline focus of MiniMax M2.1 is multi-language programming. MiniMax says it systematically improved performance across Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, and JavaScript, spanning low-level systems code through web and mobile application development. It is positioned as a 'digital employee' for end-to-end coding and agentic workflows, building on MiniMax's interleaved-thinking approach for multi-step problem solving, precise refactoring, and tool use.

On benchmarks, MiniMax M2.1 reaches 74% on SWE-bench Verified and 72.5% on SWE-bench Multilingual, which MiniMax reports as competitive with or ahead of Claude Sonnet 4.5 on coding-heavy tasks, at a fraction of the cost. With a roughly 205K-token context window, text-only inputs, and hosted pricing of $0.30 per million input tokens and $1.20 per million output tokens, M2.1 targets developers who want strong agentic coding without frontier-model API bills.

// Benchmarks

1. SWE-bench Verified74%
2. SWE-bench Multilingual72.5%
3. Multi-SWE-bench49.4%
4. Terminal-Bench 2.047.9%
5. VIBE (average)88.6%
6. Toolathlon43.5%
7. BrowseComp47.4%
8. AIME 202583%
9. MMLU-Pro88%
10. GPQA Diamond83%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Pricing source ↗

// Strengths

• Strong real-world software-engineering performance: 74% on SWE-bench Verified, competitive with leading proprietary coding models
• Broad multi-language programming coverage tuned for Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, and JavaScript
• Open weights under a modified MIT license, enabling local and self-hosted deployment via SGLang, vLLM, Transformers, and KTransformers
• Efficient sparse MoE design (only ~10B of ~230B parameters active per token) keeps inference fast and inexpensive
• Low hosted pricing at $0.30 / $1.20 per million input/output tokens, far below frontier proprietary coding models
• Built for agentic, multi-step tasks with interleaved thinking and tool use

// Best for

• Autonomous and assisted software engineering across many programming languages
• Agentic coding workflows that plan, edit, run tools, and iterate over multi-step tasks
• Code refactoring and migration in large, multi-language codebases
• Web and mobile (Android/iOS) application development assistance
• Self-hosted or on-premise coding assistants where open weights and data control matter
• Cost-sensitive automation: documentation, data handling, and multi-step pipelines at low per-token prices

// How to access

// MiniMax M-Series — every version

The full lineage of the MiniMax M-Series line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.