Every few months a Chinese lab drops a model that makes Silicon Valley wince. The June 2025 release from MiniMax (稀宇科技) was different: not because it was the biggest, but because it was cheap to make. The company says it trained a frontier-class reasoning model for about the cost of a mid-range apartment.
A different kind of open-weight bet
MiniMax, a Shanghai-based lab among China's so-called "AI tigers," kicked off a five-day "MiniMax Week" on 16 June 2025 by open-sourcing MiniMax-M1. The company describes it as the world's first open-source, large-scale hybrid-attention reasoning model — a claim carried by Chinese state and financial outlets, not an independent title we can verify, but accurate in describing what the weights actually are.
The move matters because of where the model sits. Most open-weight releases in 2025 chased raw scale. MiniMax aimed at the agentic sweet spot: long context, tool use, and reasoning that stays affordable to run. Weights landed on Hugging Face and GitHub the same day, under a license developers can actually deploy.
The architecture that made it cheap
M1 is a mixture-of-experts (混合专家) model with 456 billion total parameters but only 45.9 billion active per token. The trick is the attention layer. Instead of pure softmax attention, MiniMax bolts a linear "Lightning Attention" mechanism alongside it — seven lightning layers and one softmax layer per eight-layer block. Linear attention scales nearly linearly with sequence length instead of quadratically, so million-token inputs stop being a bankruptcy event.
On top of that sits CISPO, a reinforcement-learning algorithm that clips importance-sampling weights rather than editing token updates. MiniMax argues CISPO stabilized long reasoning chains and converged roughly twice as fast as rival methods like DAPO. That is the lab's own benchmark claim; we treat it as a description of their method, not a settled ranking.
The headline spec is the context window: 1 million tokens of input and 80,000 tokens of output. That is eight times the 128k context of DeepSeek R1, and matches the input length of Google's Gemini 2.5 Pro at the time.
The cost nobody else matched
Here is the number that traveled furthest. MiniMax says the entire reinforcement-learning stage of M1 ran on 512 Nvidia H800 GPUs for three weeks at a cost of about $535,000 (roughly HK$4.2 million). Multiple Chinese reports converge on the figure — Shanghai Securities News cites $535,000, a Hebei science-and-technology notice cites $537,400, and Hong Kong tech outlet unwire.hk cites $534,700.
For context, that is a fraction of what frontier labs spend on a single training run. And the efficiency shows at inference too: MiniMax reports that at 80,000-token deep reasoning, M1 needs only about 30% of the compute of DeepSeek R1, and at 100,000 tokens about 25%. Those ratios come from the company's technical report; they are plausible given the linear-attention design but should be read as vendor-measured.
On the Artificial Analysis Intelligence Index, M1 landed at second among open models at release — a strong, externally tracked placement that gives the cost story some backing.
An ecosystem play, not just a model
MiniMax did not ship M1 alone. "MiniMax Week" rolled out a stack, not a single drop:
- MiniMax-M1 — the open-weight reasoning base model with the 1M-token window.
- Hailuo-02 — a video model the lab says reached second on the Artificial Analysis video arena, ahead of Google Veo and Kling.
- MiniMax Agent — a general agent for long-horizon tasks, pitched as the "reliable" productivity layer.
The pattern is consistent: open the base model to pull developers in, then monetize through API, agents, and media generation.
The lab is also courting the open infrastructure world directly. MiniMax said it is working with China's National Supercomputing Center and with vLLM, the popular inference framework, so developers can deploy M1 without reinventing the serving stack. That is the quiet part of the strategy — owning the default deployment path is worth more than a one-time benchmark win.
MiniMax is reportedly backed by Tencent and Alibaba, which matters less for the model and more for the war chest behind a multi-product roadmap.
Honest limitations
- The "world's first hybrid-attention open reasoning model" label is MiniMax's own phrasing, repeated by Chinese media; we found no independent body that crowns it.
- The $535,000 figure covers only the RL stage, not the full pretraining of the underlying MiniMax-Text-01 base; the base training cost is a separate, larger number MiniMax has not put in a single headline figure.
- The 30%/25% compute-vs-DeepSeek-R1 ratios are vendor-reported and depend on MiniMax's own measurement setup.
- Hybrid linear attention is still newer than standard transformers; long-term quality on tasks beyond the benchmark set is unproven in production at scale.
What readers can do now
- If you build agents that chew long documents, pull M1 from Hugging Face and benchmark it against DeepSeek R1 on your 100k-token workloads — the cost gap is the real differentiator.
- Watch whether
vLLMand the National Supercomputing Center integrations actually ship stable serving, since that determines if the "cheap to run" promise holds outside MiniMax's cloud. - Track MiniMax's next moves in video (Hailuo) and agents as the signal of whether an open base model can fund a multi-product company.