Xiaomi gives away the second tier of AI

I.A phone maker shipped a trillion-parameter agent for free

On 22 September, Xiaomi released the MiMo-V2.6 series: a Pro model, a Flash model and a 9-billion-parameter distilled version, with weights on Hugging Face. VentureBeat reported the specifications: Pro is a sparse mixture-of-experts model with 1.02 trillion total and 42 billion active parameters; Flash has 310 billion and 15 billion. Both take a one-million-token context and carry the MIT licence, which permits commercial use and modification.

The API prices are $0.435 and $0.87 per million input and output tokens for Pro, and $0.14 and $0.28 for Flash. Xiaomi says the series keeps V2.5's pricing and costs “1/20 to 1/60” of overseas models at the same intelligence level.

Xiaomi also published what the final training stage cost: about $2.62 million of reinforcement learning for Pro and $850,000 for Flash, each run covering roughly 750,000 trajectories in under six days. That figure covers the last RL stage only, not pre-training, and it should not be read as the cost of building the model.

The headline claim is that the world's best open model now matches a closed one. That is true, and the closed model it matches is not the frontier.

II.The tie is with the second tier, not the top

Artificial Analysis scores MiMo-V2.6-Pro at 46 on its Intelligence Index. Unite.AI's roundup of the leaderboard puts that level with xAI's Grok 4.7 and ahead of the strongest other open models, GLM-5.3 at 45 and Kimi K3 at 44.

The closed frontier sits higher. On 2 October The Next Web reported the same index with Claude Opus 5.5 at 58, Gemini 4 Argon and GPT-6 Astra at 53, and GPT-6.1 Sol at 52. Xiaomi's own announcement concedes a gap to the strongest closed models. The open-weight lead is real, but it is a lead of one or two points over other open models, and it trails the frontier by twelve.

The open leaderboard also turns over fast. In May, Xiaomi was already saying its previous model, V2.5-Pro, ranked first among open models on Artificial Analysis. Being first among open weights is a position that changes hands every few months between Chinese labs, not a lasting advantage.

III.The honest technical read

The agentic-coding numbers are strong and mostly vendor-run, and the sources disagree on them. Xiaomi's announcement says Pro rose on DeepSWE v1.1 from 58.4 to 72.6 during the final RL run. VentureBeat reports 71.9. Unite.AI reports 72.57 and compares it with 19.0 for the older V2.5-Pro. These describe two different baselines: one is the same model before its last RL run, the other is the previous generation. On Unite.AI's table, Claude Opus 5 and GPT-6 Astra both score 74.0, and DeepSeek V4.1 Flash, another open model, scores 74.2. So Pro is not the open leader on this test.

Independent hands-on tests found weaker results. MindStudio put Pro at 34.9 on Terminal Bench 4.0 against 49.0 for Claude Opus 5, and 47.9 on ExploitBench against 70.0. One debugging task took more than 25 minutes. ComputingForGeeks ran infrastructure tasks on 25 September: Pro produced a working Kubernetes deployment in one of three runs, and its median wall time was 88 seconds against 18 for GPT-6 Sol. One request hung for more than ten minutes.

Latency is a real constraint. Tabbit measured about 18 seconds to first token on a 10,000-token input, most of it spent thinking. That is acceptable for overnight agent jobs and poor for interactive chat. Reported output speed also varies by source, from about 40 tokens per second on Artificial Analysis to about 130 on Unite.AI's figures.

Self-hosting is not cheap either. ComputingForGeeks lists Pro's weights at 566 GB, served across eight GPUs with tensor parallelism; Flash needs 173 GB and at least two 80 GB GPUs. “Free weights” means no licence fee. It does not mean free to run.

IV.The squeeze lands on the $2/$10 middle

Prices. Closed mid-tier models now cluster at $2/$10 per million tokens for Claude Sonnet 5.5 and GPT-6.1 Sol, with Opus 5.5 at $4/$20. MiMo-V2.6-Pro's output tokens cost about one-eleventh of the mid-tier rate and one twenty-third of Opus. Per indexed task the gap is about 25 times against GPT-6 Astra, though for a lower score. The model competes with Sonnet-class pricing. It does not compete with Opus-class capability.

Market structure. In our reading, most enterprise agent work runs on mid-tier models: classification, extraction, routine code changes and batch pipelines. A free model that matches Grok 4.7 lowers the ceiling on what closed labs can charge for that tier, while the frontier keeps its premium. Expect closed labs to defend the middle with cheaper SKUs and longer introductory pricing, not to lose it overnight.

Build versus buy. Our estimate, not a sourced figure: assume eight 80 GB GPUs rented at $2.50 per GPU-hour. Self-hosting Pro then costs about $14,600 a month. At $0.87 per million, the same money buys roughly 16.8 billion output tokens from Xiaomi's API. Below that volume, the API is cheaper. Self-hosting makes sense for very heavy users, or for firms that need data to stay on their own hardware.

Capital. Xiaomi is paying for this out of hardware margins. Lei Jun committed more than 60 billion yuan to AI over three years, including more than 16 billion yuan in 2026, and Xiaomi's shares fell more than 6% on the day. In the first quarter, revenue fell 10.9% to 99.1 billion yuan and adjusted net profit fell 43.1%. A company with shrinking profit is giving away a model to win developers for its phones, cars and its own API, not to earn money from the model itself.

Labour. The tasks a cheap 46-level model handles well are high-volume, verifiable and tolerant of delay. That puts pressure on outsourced routine development and data work first. Senior engineering is less exposed, because that work still depends on frontier models and fast feedback.

V.Second-order: open tokens, closed revenue

Open models already carry most of the traffic without earning most of the money. Open-weight models account for about 60% of OpenRouter's US token consumption, most of them Chinese, according to The New Stack on 14 September. But a Linux Foundation analysis of OpenRouter traffic, cited by Tech Times, found closed providers took about 96% of model-layer revenue from May to September 2025. That data is a year old. It is still the strongest evidence against the idea that cheaper open models drain closed-lab revenue.

Procurement is the bottleneck. The same New Stack report cites Deloitte finding that 77% of companies weigh a vendor's country of origin. OpenRouter now offers US-only routing for Chinese models hosted by US providers. A Chinese-built model only reaches regulated buyers after a Western host runs it, which hands part of the margin to the infrastructure layer.

The release also lowers costs for other labs. Xiaomi says it will open-source more than 7,000 RL task environments and its training framework. If that happens, the cost of post-training an agent model falls for every competitor, open or closed. And Xiaomi's 9B distilled model is a cheaper path to fine-tuned agents at the edge, though ComputingForGeeks found it failed most of its infrastructure tasks.

VI.Precedent: the DeepSeek shock, with less noise

In January 2025, DeepSeek's R1 release coincided with Nvidia losing about $589 billion in market value in one day. The fear was that cheap training would cut demand for chips.

We found no report of a comparable market reaction to MiMo-V2.6, and that is the point. Investors now expect a cheap Chinese open model a few points behind the frontier every few months. The effect shows up slowly, in mid-tier price lists and enterprise model routing, rather than in a single day's share price.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not forecasts from a model, and they sum to 100%.