The release notes landed on a Monday, and within days the weight files were on Hugging Face for anyone to download. MiniMax, the Chinese AI company best known for its Hailuo video models and its talkative companion app, announced M3 on June 1, a model it bills as the first open-weights release to combine frontier-level coding ability with agentic behavior in a single artifact. Developers who had been stitching together a reasoning model, an agent harness and a long-context model to build autonomous coding tools suddenly had an alternative that does all three at once.
The specifications explain the interest. M3 is a mixture-of-experts model with 428 billion total parameters, activating about 23 billion per token. It supports a 1 million-token context window, enough to hold an entire large codebase, and handles text, images and video natively. The engine underneath is a new attention design MiniMax calls MSA, short for MiniMax Sparse Attention, which the company says cuts per-token compute at long contexts to about one-twentieth of its predecessor M2 and delivers a 15-fold decode speedup at 1 million tokens. On SWE-Bench Pro, a benchmark that measures real-world software engineering, MiniMax claims a score of 59 percent, a figure that places it just behind far more expensive closed models.
The pricing is where M3 gets attention. The company lists the model at $0.30 per million input tokens and $1.20 per million output tokens. Compare that with leading closed frontier models, which charge around $25 per million output tokens, and the gap is roughly twenty-fold. For teams running heavy agentic workloads, where a single autonomous coding session can consume millions of tokens, the arithmetic is hard to ignore. The model is available through MiniMax’s own API and platform, and the weights were committed for release within roughly ten days of the announcement.
MiniMax is careful about what it claims. The benchmark scores are vendor-run, and the company has not disclosed every training detail, which has prompted some researchers to caution that third-party verification is still needed. But the release has drawn attention from the open-source community because of what it represents: a Chinese lab shipping open weights that compete with closed frontier models on coding and agent tasks while undercutting them on price. MiniMax’s engineering head, who goes by SkylerMiao on social media, has cross-confirmed the framing that M3 is the first open-weights model to fuse reasoning and agent behavior in one release.
The timing matters. Agentic coding, where a model plans, writes code, runs tools and iterates on failures, has become the most competitive segment in AI, with Microsoft, OpenAI, Anthropic and Google all pushing agents into developer workflows. MiniMax’s bet is that a significant share of that work will happen with open weights, either self-hosted or through cheaper APIs, and that the 1 million-token context window gives it an edge for teams working on large repositories. The company has been building toward this position for years, shipping the M2 line of open-weight models and the Hailuo video series, and M3 is its most aggressive attempt yet to convert that base into a frontier claim.
There are real questions about how far open weights can go. Closed models still lead on many benchmarks, and the operational burden of running a 428-billion-parameter model, even with sparse attention, is not trivial. Quantized versions have been demonstrated running locally on a 512-gigabyte Apple workstation, but most teams will rent compute rather than buy it. The company’s answer is that open weights plus low API prices give developers optionality, and that the community will do what it does best: fine-tune, adapt and deploy the model in places a single vendor would never reach.
For the broader market, M3 is another sign that the frontier of open-weight AI is moving from research curiosities toward production infrastructure. The pattern is familiar to anyone who watched Linux displace proprietary operating systems, or open-source databases erode the market share of commercial ones. A model that matches most frontier capabilities at a fraction of the price, with weights anyone can inspect and modify, changes the negotiating position of every enterprise evaluating AI vendors. Analysts who track the sector say the competitive pressure is already visible in closed-model pricing, which has fallen sharply over the past two years.
MiniMax itself is playing a longer game. The company, valued at well over $3 billion after a funding round in 2025, has spent heavily on compute and talent to chase the frontier, and M3 is the product of that investment. The bet is that agentic AI becomes the default way people use models, and that the winners will be the labs that can combine capable models, cheap inference and developer trust. M3 does not settle that contest. But it gives the open-weight camp its strongest candidate yet, and it gives developers a reason to question why they should pay twenty times as much for a closed alternative.
