one year on
MiniMax open-sources M2, a 230B MoE model tuned for coding agents, priced at 8% of Claude Sonnet 4.5
Shanghai's MiniMax releases MiniMax-M2 weights on Hugging Face and undercuts Anthropic's Sonnet 4.5 on price and speed, and says cheap, fast agents are the target.
Originally published Published on this site
MiniMax released and open-sourced MiniMax-M2 today, a mixture-of-experts model with 230 billion total parameters and 10 billion activated per token, built specifically for coding and agentic workflows. The Shanghai lab published the weights on Hugging Face, with deployment guides for vLLM and SGLang, and is offering the hosted API free for a limited time — a window it says is now extended to November 7th at 00:00 UTC.
The pitch is cost and speed. MiniMax lists API pricing at $0.30 per million input tokens and $1.20 per million output tokens, and says that is 8% of Claude Sonnet 4.5’s price with roughly double the inference speed (around 100 tokens per second). M2 emphasizes end-to-end developer workflows — multi-file edits, coding-run-fix loops and test-validated repairs — and names Claude Code, Cursor, Cline, Kilo Code and Droid as target environments. Benchmark numbers put it at 69.4 on SWE-bench Verified, and MiniMax says its tool-use and deep-search abilities are very close to the best overseas models. On the Artificial Analysis benchmark, it says M2 ranks in the top five globally.
The record
Positioned M2 as closing what it called the 'impossible triangle' of performance, price and inference speed, arguing existing agent products are 'very expensive or slow to achieve good results.' MiniMax said it extended M2's free trial to November 7th, 00:00 UTC after what it described as extensive testing by developers at home and abroad over the past weekend.
Framed the release as double the inference speed of Anthropic's Claude Sonnet 4.5 at 8% of its API cost, with weights on Hugging Face and vLLM/SGLang support.
One year later — open only if you can handle spoilers
M2 set MiniMax's pace: the company more than doubled on its Hong Kong IPO debut on January 9, 2026, and followed with M2.5, M2.7 and, in June 2026, the open-weight M3. In February 2026 Anthropic said MiniMax had run more than 13 million exchanges with Claude through fraudulent accounts to distill its agentic-coding abilities, part of a campaign it also attributed to DeepSeek and Moonshot.
Filed
Compiled and edited by Adam Allcock.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy