MiniMax M2.1

MiniMax’s advanced coding-focused language model released December 23, 2025.

Key Specifications

• Release Date: December 23, 2025
• Architecture: Sparse Mixture-of-Experts (MoE)
• Total Parameters: 230 billion
• Active Parameters: 10 billion per token
• Context Window: 200,000 tokens
• Quantization: FP8 native
• Attention: Multi-Head Attention (MHA)
• Modalities: Text, image, audio, video inputs

Pricing

• Input: $0.30 per million tokens
• Output: $1.20 per million tokens

Benchmarks

Benchmark	M2.1 Score	Notes
SWE-Multilingual	72.5%	Outperforms Claude Sonnet 4.5, Gemini 3 Pro
VIBE (Full-Stack)	88.6%	Web: 91.5, Android: 89.7
SWE-bench Verified	74.0%	Up from M2’s 69.4%
Multi-SWE-bench	49.4%	Up from M2’s 36.2%
MMLU	88	General knowledge

Key Capabilities

Coding Focus

• Multi-file code editing
• Code-run-fix loops
• Multi-language programming (Python, JS, Go, Rust, etc.)
• Full-stack web and mobile development
• Native Android and iOS development
• Godot game engine and C++ graphics

Efficiency

• Only 10B active parameters = low latency
• Runs on consumer H100 or dual RTX 4090
• Concise, high-efficiency responses
• Near-instant latency for developer workflows

Long Context

• 200K tokens (up from M2’s 128K)
• Process entire product specs, design docs
• Simplified RAG setups

Improvements Over M2

• Context: 128K → 200K tokens
• SWE-Multilingual: 56.5% → 72.5%
• Multi-SWE-bench: 36.2% → 49.4%
• Advanced Interleaved Thinking for reasoning
• Reduced latency and verbosity

Use Cases

• AI coding assistants (Claude Code, Cline, Roo Code, Droid)
• Long-horizon developer toolchains
• Complex agentic workflows
• Full-stack application development

See Also

• MiniMax M1 - Earlier open-weight model
• MiniMax-Text-01 - Foundation text model