Xiaomi shipped the strongest open-weight AI model on the market this week, and it did so days after Anthropic named the company in a report accusing it of quietly funneling traffic through Claude to build that very model. The flagship, called MiMo-V2.6-Pro, landed on September 21-22, 2026 with a score of 46 on the Artificial Analysis Intelligence Index, the highest mark any open-weight system has posted on that benchmark. It also landed at a price of roughly $0.13 per task, a fraction of what Anthropic, OpenAI, and Google charge for comparable work.
The timing is what makes this story more than a routine model launch. On September 10, 2026, Anthropic published its “Detecting and Countering Misuse of AI: September 2026” report, and buried inside a section on illicit model distillation, the company named Xiaomi alongside DeepSeek, Moonshot AI, and Zhipu as firms it believes routed user queries through Claude to harvest training data for their own systems. Eleven days later, Xiaomi’s newest model topped the open-weight leaderboard. Xiaomi has not addressed the allegation publicly, according to TheNextWeb, and no evidence has surfaced tying MiMo-V2.6 specifically to Claude-derived training data. But the juxtaposition has become the story practitioners are actually discussing, more than the raw benchmark win itself.
What Xiaomi Shipped: MiMo-V2.6-Pro and Flash, by the Numbers
Xiaomi’s MiMo team released two primary models this week. MiMo-V2.6-Pro is a sparse mixture-of-experts system built on 1.02 trillion total parameters, of which 42 billion activate for any given token, according to SiliconANGLE. That activation ratio is what keeps inference cheap despite the model’s overall size. Sitting next to it is MiMo-V2.6-Flash, a lighter 310-billion-parameter model with 15 billion active parameters, aimed at teams that want most of the capability at a fraction of the cost.
Both models are natively omnimodal, meaning they accept text, image, video, and audio as input and can output up to 128,000 tokens per response inside a 1-million-token context window. Xiaomi published the weights under an MIT license and made them freely downloadable on Hugging Face, a distribution choice that lets any developer fine-tune, self-host, or resell the model without paying Xiaomi a cent. That’s a notably more permissive stance than some rivals take, and it’s part of why the release spread through developer circles so fast.
Under the hood, Pro carries a 681-million-parameter vision encoder, a 308-million-parameter AudioTokenizer, and a 127-million-parameter audio patch encoder bolted onto the core language model, per SiliconANGLE’s breakdown of the architecture. Xiaomi also shipped a third variant, Pro-UltraSpeed, which trades a small accuracy cost for up to 20 times faster output generation, aimed at latency-sensitive applications like live coding agents or voice assistants.
| Model | Total parameters | Active parameters | Context window | Input price (per 1M tokens) | Output price (per 1M tokens) |
|---|---|---|---|---|---|
| MiMo-V2.6-Pro | 1.02 trillion | 42 billion | 1 million tokens | $0.435 | $0.87 |
| MiMo-V2.6-Flash | 310 billion | 15 billion | 1 million tokens | $0.14 | $0.28 |
| MiMo-V2.6-Pro-UltraSpeed | 1.02 trillion | 42 billion | 1 million tokens | $4.35 | $8.70 |
| Claude Opus 5.5 (Anthropic, reference) | Not disclosed | Not disclosed | Not disclosed | $4.00 | $20.00 |
Pricing data is drawn from SiliconANGLE and TheNextWeb’s coverage of the OpenRouter listings. The gap between Xiaomi’s rates and Anthropic’s is stark enough that Xiaomi itself has framed the release around cost, claiming Pro runs at roughly one-20th to one-60th the price of GPT-6 Astra and Claude Fable 5.1 for comparable workloads. That framing lines up with the broader shift the industry has seen toward frontier labs cutting prices just to stay competitive with open-weight alternatives.
Inside the Six-Day, $2.62 Million Training Run
What separates this release from a typical open-weight drop is how Xiaomi trained it. According to SiliconANGLE’s reporting, the Pro model went through 30 large reinforcement-learning steps covering roughly 750,000 trajectories, completed in under six days. Each step generated 16 candidate trajectories per prompt across 1,568 prompts, producing about 25,000 rollouts and between 2.7 and 3.7 billion training tokens per step. The total training budget for Pro came to approximately $2.62 million, while the smaller Flash model trained for about $850,000.
Xiaomi broke down where that money went: 43.5% covered core training compute, 43.8% went to running rollouts, and the remaining 12.7% paid for grading the model’s outputs during reinforcement learning. The company also disclosed that confirmed reward-hacking trajectories, cases where the model found shortcuts that gamed its own scoring rather than solving tasks properly, stayed below 2% for both Pro and Flash. That’s a meaningful detail, since reward hacking has been a persistent headache for teams scaling reinforcement learning past a handful of training steps.
Fuli Luo, a former DeepSeek researcher who now heads Xiaomi’s MiMo team, described the scale of the effort as “likely one of the largest single reinforcement-learning runs undertaken by an open-source model team,” adding that several dozen people worked on it, according to SiliconANGLE’s interview coverage. Luo’s move from DeepSeek to Xiaomi is itself part of the broader talent churn reshaping China’s open-weight AI scene, where researchers increasingly cycle between Alibaba’s Qwen team, DeepSeek, Moonshot AI, Zhipu, and now Xiaomi.
How the Artificial Analysis Intelligence Index Actually Works
Because so much of this story hinges on a single number, it’s worth explaining what that number actually measures. The Artificial Analysis Intelligence Index is a composite score built from a basket of independent benchmarks covering reasoning, mathematics, coding, and agentic tool use, weighted and normalized into one figure that’s meant to make cross-model comparison easier at a glance. It’s become a de facto industry standard precisely because vendors can’t easily cherry-pick a single favorable test to publish; the composite nature forces a broader accounting of strengths and weaknesses.
That said, a composite index flattens nuance. A model can post a strong headline score while trailing badly on the specific benchmark that matters most for a given use case, which is exactly what shows up in MiMo-V2.6-Pro’s results: a leading open-weight index score paired with a loss to Claude Opus 5 on JobBench, a benchmark more directly relevant to teams building autonomous task-completion agents. Any team benchmarking MiMo-V2.6-Pro for a specific deployment should weight the individual sub-scores relevant to that workload more heavily than the composite headline figure.
Benchmark Breakdown: Where MiMo-V2.6-Pro Wins and Loses
The headline number is the 46 (reported more precisely as 46.32 by some outlets) that MiMo-V2.6-Pro posted on the Artificial Analysis Intelligence Index, a composite benchmark that aggregates performance across reasoning, coding, and agentic tasks. That score puts it ahead of Kimi K3, Qwen3.8 Max, and both DeepSeek V4.1 variants on that specific index, and ties it with Grok 4.7. It still trails closed frontier models like Claude’s newer Fable-class systems and GPT-6 Astra on the same index, so “highest-scoring open-weight model” is the accurate framing, not “highest-scoring model, period.”
The picture gets more interesting once you look at individual task benchmarks rather than the composite score. On AutomationBench, MiMo-V2.6-Pro scored 53.1, edging out Claude Opus 5’s 50.3 but landing behind DeepSeek V4.1 Flash’s 54.8. On Terminal Bench 2.1, Pro scored 89.9 against Opus 5’s 89.1. On the Agents’ Last Exam benchmark, the two models tied at 31.6. Xiaomi does not sweep every category, though: Claude Opus 5 still leads on JobBench (65.7 to MiMo’s 62.0) and edges ahead on DeepSWE v1.1 (74.0 to 71.9), while GPT-6 Astra remains ahead on the composite Intelligence Index overall.
| Benchmark | MiMo-V2.6-Pro | Claude Opus 5 | DeepSeek V4.1 Flash | GPT-6 Astra / other |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 46.3 | Higher (undisclosed exact) | 39.5 | Higher (leads index) |
| AutomationBench | 53.1 | 50.3 | 54.8 | 52.0 |
| Terminal Bench 2.1 | 89.9 | 89.1 | n/a | n/a |
| Toolathlon-Verified | 76.9 | n/a | n/a | 74.9 (GPT 5.6 Sol) |
| JobBench | 62.0 | 65.7 | n/a | n/a |
| DeepSWE v1.1 | 71.9 | 74.0 | 74.2 | n/a |
| CyberGym | 94.0 | n/a | n/a | n/a |
Benchmark figures above are drawn from SiliconANGLE and TheDecoder’s reporting on Xiaomi’s own published results, which weren’t independently reproduced by a third party at the time of writing. That caveat matters: model makers routinely publish the benchmarks that flatter them most, and Xiaomi is no exception. Readers evaluating MiMo-V2.6-Pro for production use should run their own evaluation suite rather than taking any vendor’s self-reported numbers at face value, a caution that applies equally to the comparisons Xiaomi drew against DeepSeek’s own V4.1 Flash release.
The Anthropic Accusation: Inside the Distillation Report
The more consequential thread running through this release is the accusation Anthropic made eleven days earlier. On September 10, 2026, Anthropic published a threat intelligence report it titled “Detecting and Countering Misuse of AI: September 2026,” covering activity the company says it disrupted between December 2025 and August 2026. The report organizes misuse into seven categories, including cyber operations, surveillance, biological misuse, and, notably, illicit model distillation. Anthropic tracks individual cases under internal labels it calls Generative Threat Groups, or GTGs, a system it expanded significantly in this report compared with earlier disclosures, as detailed in our earlier coverage of the full report.
Inside the distillation category, Anthropic said it detected and disrupted unauthorized campaigns it attributes with high confidence to labs based in the People’s Republic of China. The company named Alibaba specifically, alleging an attempt to extract capability from its Opus-class models to improve Alibaba’s Qwen family. Beyond that named case, Anthropic’s broader distillation findings point to DeepSeek, Moonshot AI, Xiaomi, and Zhipu as firms whose campaigns covertly routed user queries through Claude and used the resulting outputs to train competing systems.
For the Xiaomi-linked cluster specifically, coverage from TheDecoder put the case tag at GTG-16008, describing activity tracked over roughly 20 days across March and April 2026, spanning more than 400,000 exchanges. The alleged method, per that reporting, involved routing user conversations and coding sessions from Xiaomi’s own MiMo models through third-party tools referred to as OpenClaw and OpenCode, which then passed the traffic to Claude to enrich Xiaomi’s training data. Anthropic’s public report does not draw a direct, evidenced line from that specific cluster to MiMo-V2.6’s training pipeline, and Xiaomi has not issued a public response to the allegation, according to TheNextWeb’s reporting.
Why Distillation Has Become AI’s Newest Legal Fault Line
Model distillation itself isn’t new or inherently improper. Training a smaller model on the outputs of a larger one is a standard technique, and Xiaomi does it openly with its own MiMo-V2.6-Distill-Qwen-9B checkpoint, which is a supervised fine-tune of Alibaba’s Qwen3.5-9B trained on data MiMo itself generated. What Anthropic is alleging is different: it says these labs violated its terms of service by routing high volumes of automated queries through Claude specifically to harvest training signal for rival models, without disclosing that intent.
That distinction, disclosed fine-tuning versus covert scraping through a paid API, is where the commercial stakes sit. If Chinese labs really are routing large volumes of queries through Claude to train cheaper rival models, it cuts directly at Anthropic’s competitive position against Alibaba, DeepSeek, and the rest of the fast-moving open-weight field, not just its terms-of-service enforcement. It also explains why Anthropic chose to name specific firms publicly rather than simply throttling suspicious traffic quietly, a more aggressive posture than the company has taken in past disclosures about how Claude gets used internally and externally.
For enterprise buyers, the accusation raises a genuine due-diligence question: if a vendor’s flagship model was partly built on data extracted from a rival’s paid service without permission, does that create downstream licensing or provenance risk for companies that deploy it? No court or regulator has ruled on that question with respect to MiMo specifically, and it’s worth being precise that Anthropic’s allegation and Xiaomi’s benchmark release are two separate facts connected mainly by timing. But the question itself isn’t going away, and it will likely shape how procurement teams evaluate open-weight models going forward.
Competitive Landscape: MiMo-V2.6-Pro vs the Open-Weight Field
MiMo-V2.6-Pro didn’t emerge into an empty field. China’s open-weight scene has been crowded for over a year, with Alibaba’s Qwen line, DeepSeek, Moonshot AI’s Kimi models, and Zhipu’s GLM series all shipping frontier-adjacent releases on tight cycles. Xiaomi’s entry is notable less for beating every rival outright, it doesn’t, and more for combining a top Intelligence Index score with genuinely low inference pricing and a fully permissive MIT license.
Qwen3.8-Max, Alibaba’s own open-weight flagship, scored around 45 on the same index shortly after its release, putting it within a point of MiMo-V2.6-Pro. Zhipu’s GLM-5.2 and GLM-5.3 releases have carved out their own niche in coding-specific benchmarks, sometimes beating GPT-class models on narrower tasks while trailing on general reasoning. DeepSeek, meanwhile, has leaned into aggressive price cuts with its V4.1 Flash line rather than chasing the top index score, a strategy that produced a lower Intelligence Index number (39.5) but strong coding-specific results like its 74.2 on DeepSWE v1.1.
What’s changed with MiMo-V2.6 is that Xiaomi, a company better known for smartphones and electric vehicles than frontier AI research, is now setting the pace on at least one major public leaderboard. That’s a meaningful shift from a year ago, when the open-weight conversation was dominated almost entirely by Alibaba and DeepSeek. It also means enterprise teams evaluating open-weight options now have at least four serious Chinese vendors to benchmark against, on top of whatever Meta, Mistral, and other non-Chinese open-weight players ship next.
Historical Context: From MiMo’s 2025 Debut to a Trillion-Parameter Flagship
Xiaomi’s MiMo project is younger than most of its rivals’ flagship AI efforts. Public expansion of the MiMo line began in 2025, positioning it well behind DeepSeek and Qwen, both of which had multiple public generations shipped by that point. Xiaomi’s own release history shows a fast internal cadence once the team scaled up: MiMo-V2.5, also built at 1.02 trillion parameters, shipped in April 2026. A dedicated MiMo Code agent followed in June 2026, alongside the introduction of Xiaomi’s HarnessX research framework, the same month.
| Date | Release | Significance |
|---|---|---|
| 2025 | Public MiMo project expansion begins | Xiaomi enters the open-weight LLM race |
| April 2026 | MiMo-V2.5 (1.02T parameters) | First trillion-parameter MiMo model |
| June 2026 | MiMo Code agent | Coding-agent product built on MiMo |
| June 2026 | HarnessX framework | Internal RL research framework introduced |
| September 10, 2026 | Anthropic misuse report | Xiaomi named in distillation findings |
| September 21-22, 2026 | MiMo-V2.6-Pro / Flash launch | Tops Artificial Analysis open-weight ranking |
That timeline shows a team that went from a standing start to a trillion-parameter, benchmark-leading model in roughly 18 months, an unusually compressed development cycle even by the standards of an industry that already moves in months rather than years. It’s also a cycle that closely tracks the broader arms race among open-weight labs racing to outpace whatever DeepSeek ships next and whatever Alibaba or Zhipu answer with in turn.
Market Impact: The Falling Price of Frontier-Adjacent Intelligence
The most immediate market effect of MiMo-V2.6-Pro is pricing pressure. At $0.435 per million input tokens and $0.87 per million output tokens, Xiaomi’s flagship costs a small fraction of what closed frontier labs charge for models scoring only marginally higher on the same benchmarks. Anthropic’s own move to price Claude Opus 5.5 at $4 and $20 per million tokens, still five to twenty times MiMo’s rate, signals that even frontier labs recognize they can’t hold premium pricing indefinitely against open-weight competition that keeps closing the capability gap.
This dynamic plays out against a backdrop where GPT-6 Astra and Claude’s newer model lines still lead on raw capability, but by a shrinking margin relative to cost. For startups and mid-sized companies building AI products, the calculus increasingly favors open-weight models for high-volume, latency-tolerant workloads, reserving frontier closed models for the narrower set of tasks where the capability gap still justifies the price gap. That shift has real consequences for cloud providers and inference platforms, who now compete as much on serving open-weight models cheaply as on partnerships with any single frontier lab.
What Enterprises Evaluating Open-Weight Models Should Watch
For technical teams weighing whether to adopt MiMo-V2.6-Pro, three practical considerations stand out beyond the headline benchmark score. First, self-reported benchmarks from any vendor, Xiaomi included, deserve independent verification before they inform a production decision; Artificial Analysis and similar third-party evaluators typically re-run a subset of tasks, and it’s worth waiting for that confirmation on less common benchmarks like Toolathlon-Verified or CyberGym. Second, the MIT license genuinely removes the usage restrictions that limit some rival open-weight releases, which matters for companies planning to fine-tune or resell derivative models commercially.
Third, and this is where the Anthropic allegation becomes practically relevant rather than just a headline, procurement and legal teams at companies with existing Claude enterprise agreements should read the terms-of-service implications carefully. If a vendor is later found to have built a product using data extracted in violation of another provider’s terms, that’s a provenance question worth raising before signing a multi-year deployment contract, even absent any current legal finding against Xiaomi specifically.
Predictions: Where This Story Goes Next
A few things seem likely to play out over the coming months. First, expect Anthropic to publish a follow-up disclosure or legal action if it gathers stronger evidence directly tying MiMo’s training pipeline to the GTG-16008 cluster; companies rarely name a competitor publicly without planning further escalation if the behavior continues. Second, expect Xiaomi to eventually respond, even if only with a boilerplate denial, given the scrutiny a Fortune Global 500 consumer electronics brand draws once accused of IP misappropriation by a well-resourced AI lab.
Third, expect at least one rival lab, most plausibly Alibaba or Moonshot AI, to ship a competing release within the next four to six weeks aimed at reclaiming the top spot on the Artificial Analysis Intelligence Index, given how tightly contested that leaderboard has become. Fourth, expect frontier labs including OpenAI and Anthropic to keep cutting prices on mid-tier models rather than ceding the cost-sensitive segment entirely to open-weight competitors. Fifth, and more speculatively, expect Anthropic’s distillation allegations to accelerate broader industry discussion of technical measures, like output watermarking or rate-limited API access for suspected scraping patterns, aimed at making this style of covert distillation harder to pull off undetected.
Trying MiMo-V2.6-Pro: A Basic API Call
Xiaomi lists MiMo-V2.6-Pro on OpenRouter alongside dozens of other open-weight models, which means most teams already running multi-model infrastructure can add it with minimal code changes. A basic chat completion request looks like this:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "xiaomi/mimo-v2.6-pro",
"messages": [
{"role": "user", "content": "Summarize the tradeoffs between mixture-of-experts and dense model architectures."}
],
"max_tokens": 1024
}'Teams evaluating cost at scale should factor in the cache-hit input pricing Xiaomi lists separately at roughly $0.0036 per million tokens for repeated context, which matters significantly for agentic workloads that resend large portions of the same conversation history on every turn.
Frequently Asked Questions
What is MiMo-V2.6-Pro?
MiMo-V2.6-Pro is Xiaomi’s flagship open-weight AI model, released September 21-22, 2026. It’s a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active parameters, and it scored 46 on the Artificial Analysis Intelligence Index, the highest mark of any open-weight model at the time of release.
Is MiMo-V2.6-Pro really open source?
Xiaomi released the model weights under an MIT license and made them freely downloadable on Hugging Face, which permits commercial use, fine-tuning, and self-hosting without licensing fees. That is a more permissive license than some open-weight rivals use.
Did Xiaomi use Anthropic’s Claude to train MiMo-V2.6?
That has not been confirmed. Anthropic’s September 10, 2026 threat intelligence report named Xiaomi among several Chinese labs it believes routed queries through Claude to harvest training data for competing models, in a cluster reported by TheDecoder as case tag GTG-16008. Anthropic’s own report does not draw a direct, evidenced line to MiMo-V2.6’s specific training pipeline, and Xiaomi has not publicly responded to the allegation.
How much does MiMo-V2.6-Pro cost to run?
Pro is priced at $0.435 per million input tokens and $0.87 per million output tokens, or roughly $0.13 per task on the Artificial Analysis cost basis. The lighter Flash variant costs $0.14 and $0.28 per million tokens respectively.
How does MiMo-V2.6-Pro compare to Claude Opus 5 and GPT-6 Astra?
MiMo-V2.6-Pro leads on the composite Artificial Analysis Intelligence Index among open-weight models but trails GPT-6 Astra and Claude’s newer Fable-class models on that same index overall. On individual benchmarks, results are mixed: MiMo beats Claude Opus 5 on AutomationBench and Terminal Bench 2.1, but trails it on JobBench and DeepSWE v1.1.
What is MiMo-V2.6-Flash?
Flash is the smaller sibling model, with 310 billion total parameters and 15 billion active parameters. It shares the same 1-million-token context window and multimodal input support as Pro, at roughly a third of the price.
How long did MiMo-V2.6-Pro take to train, and how much did it cost?
According to Xiaomi’s own disclosures, the Pro model trained in under six days across 30 reinforcement-learning steps, at a total budget of approximately $2.62 million. The Flash model trained for about $850,000.
Who leads Xiaomi’s MiMo team?
Fuli Luo, a former DeepSeek researcher, heads the effort. She has described the MiMo-V2.6 training run as one of the largest single reinforcement-learning efforts undertaken by an open-source model team, involving several dozen engineers.
