MiniMax dropped M2.1 today, December 23, positioning it as the strongest open-weight model for agentic coding workflows. The company released full model weights on Hugging Face and made the API available through its platform, continuing the aggressive open-source push that's defined the M2 series.
The numbers MiniMax is reporting: 72.5% on SWE-multilingual (a benchmark testing code generation across programming languages) and 88.6% on VIBE-bench, a new evaluation the company built and plans to open-source. VIBE tests full-stack app development from web to mobile to backend. MiniMax claims M2.1 beats Claude Sonnet 4.5 and Gemini 3 Pro on these metrics, though all scores are self-reported and use the company's own scaffolding. Independent verification is pending.
The model runs on a mixture-of-experts architecture: 230 billion total parameters, but only 10 billion activated per inference. That's the efficiency play. MiniMax priced the previous M2 at roughly 8% of Claude Sonnet's token cost, and M2.1 continues that positioning.
Developer tooling is already lined up. Vercel's AI Gateway, Kilo Code, Cline, and Roo Code announced same-day integrations. The model supports both native MiniMax API calls and an Anthropic-compatible endpoint for easier migration.
MiniMax, founded in 2021 and backed by Alibaba and Tencent, is preparing for a Hong Kong IPO reportedly targeting Q1 2026 at a $4 billion valuation. The company has raised over $1 billion to date.
The Bottom Line: M2.1 gives developers another open-weight option for coding agents, with benchmark claims that need third-party validation before anyone should take them at face value.
QUICK FACTS
- 10B activated parameters (230B total)
- 72.5% on SWE-multilingual (company-reported)
- 88.6% on VIBE-bench aggregate (company-reported)
- Weights available on Hugging Face under open license
- API pricing: $0.30 per million input tokens, $1.20 per million output tokens (M2 pricing; M2.1 pricing unconfirmed)
