An MIT-licensed Xiaomi model scores 46 on Artificial Analysis's Intelligence Index
Artificial Analysis measured the score itself, and the weights are on Hugging Face under MIT, but MiMo-V2.6-Pro holds 1.02 trillion parameters, so most teams will reach it through an API at $0.43 per million input tokens.
The Scientist · Science desk
What happened
- The model is a frozen-router mixture of experts with 1.02 trillion total parameters and 42 billion active, plus hybrid attention and a five-layer speculative decoder.
- Xiaomi's reported benchmark table puts Pro at 89.9% on Terminal Bench 2.1, ahead of both Claude Opus 5 and GPT-5.6 Sol.
- Measured on Xiaomi's API, Artificial Analysis clocked 129.7 output tokens a second, against a 77.4 median for comparable open-weight models.
Why it matters
- What it limits constraint Xiaomi's hosted output price is already about half the median for its size class, so a team weighing self-hosting has to clear a low per-token price with its own accelerators before the free licence saves it money.
- Who has to decide decision Teams that had ruled out open weights on capability now face a narrower question: whether they need weights they can fine-tune and run on their own hardware, or simply a cheap endpoint.
- What doesn't add up contradiction The independently run number is the aggregate index score; the agentic results beating Claude Opus 5 and GPT-5.6 Sol are Xiaomi's own, so procurement leaning on Terminal Bench is leaning on the vendor's harness.
- What's now possible capability Fine-tuning and redistributing a reasoning model that accepts speech and video input no longer requires negotiating a commercial licence.
Next story · Build · Sep 22
Xiaomi priced six days of reinforcement learning at $3.47 millionConfirmed7 publishers
About 4 percent of MiMo-V2.6-Pro's parameters fire on any given token, so the compute per generated token is modest [19]. Serving still means holding all 1.02 trillion of them, because a frozen router can select any expert [1]. A team that downloads the MIT weights takes on that hardware requirement [3]. Xiaomi's own hosted price for the same model is $0.43 per million input tokens and $0.87 per million output [23], and a self-hosted deployment has to beat that.
Artificial Analysis ranks open-weight models only against other open-weight models in the same size class, and anything above 150 billion total parameters lands in its Large bucket [14]. So the median of 18 that MiMo-V2.6-Pro clears is a peer-group median [8]. Forkast reported the score as effectively tying Grok 4.7 on version 4.3 of the index [27].
The other numbers came from Xiaomi. DeepSWE v1.1 came in at 71.9%, behind Claude Opus 5 at 74.0% [28]. CyberGym hit 94.0%, ahead of every model in the table [30]. AutomationBench v1.0.6 landed at 53.1% [31]. Forkast flagged all of them as vendor-reported and applied the standard caution about internal testing environments [32]. A vendor's agentic run shows the model completing those tasks in the vendor's harness; reliability against your tool schemas, after a failed call on turn forty, is a separate measurement.
Artificial Analysis published what its own run cost: $206.66 for the full Intelligence Index, during which the model emitted 140 million output tokens [11][25]. At the listed $0.87 per million output tokens, those tokens account for $121.80, leaving $84.86, which at $0.43 per million input implies roughly 197 million input tokens if list prices applied throughout with no cache discount [20].
The same page labels the price twice, differently. Its summary calls $0.43 input "somewhat expensive" against a median of $0.30 and $0.87 output "moderately priced" against $1.13; its FAQ calls the identical figures "better than average" against a median of $0.45 and "very competitive" against $1.68 [24][23]. Against the higher median, the output price is about 48 percent below [26]. The page also calls the model "somewhat verbose" while giving the size-class median output count as the same 140 million tokens [25].
Forkast's earlier coverage put the training dashboard at $432,000 a day [33]. That is a training budget; the inference prices are a separate one. Flash lists at $0.14 and $0.28 per million tokens, Pro at $0.435 and $0.87 [4].
The two sources disagree on the date: Artificial Analysis lists the release as September 21, 2026, Forkast as September 22 [22][21]. Forkast also argued that US export controls, including the block on H200 imports in January 2026, acted as a catalyst for Chinese domestic independence [16][34]. The figure under that argument is real enough: domestic chipmakers took 41% of China's accelerator market in 2025, up from near zero in 2022 [17]. There is no counterfactual here for what that share would have been without the controls.
Both variants are on Hugging Face under MIT, and the hosted routes are Xiaomi's open platform API, OpenRouter and AI Studio [3][18]. The model takes text, image, speech and video input and returns text [6].
What to watch
- Whether Artificial Analysis or another independent evaluator runs DeepSWE, Terminal Bench and CyberGym on Pro.
- Whether independent hosts list the open weights below Xiaomi's own $0.43 and $0.87 per million tokens, which would show the weights serve cheaper than the vendor API.
- Whether Alibaba's V900, announced for Q1 2027 production with 500,000-chip clusters, ships on that schedule.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+35
- Incentives45
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
MiMo-V2.6-Pro uses a frozen-router Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters, with hybrid attention mechanisms and a 5-layer MTP speculative decoder.
- [2]
The design maintains a 1-million-token context window with native multimodal capabilities.
- [3]
Both the Pro and Flash variants are available with open weights on Hugging Face under an MIT license.
- [4]
MiMo-V2.6-Flash is priced at $0.14/$0.28 per million tokens and the Pro version at $0.435/$0.87 per million tokens, per Forkast, which notes these are vendor-reported figures.
- [5]
MiMo-V2.6-Pro is a reasoning model that uses extended thinking or chain-of-thought reasoning before answering.
- [6]
MiMo-V2.6-Pro supports text, image, speech and video input, outputs text, and has a 1.0M token context window.
- [7]
On a blended 7:2:1 cache hit/input/output ratio, MiMo-V2.6-Pro works out to $0.18 per 1M tokens, and pricing may vary by provider.
- [8]
MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 18).
- [9]
The MiMo team is led by Luo Fuli, formerly of DeepSeek.
- [10]
Training used fully asynchronous GRPO, a method leveraging 1,568 prompts across 16 rollouts per step.
- [11]
In total it cost $206.66 to evaluate MiMo-V2.6-Pro on the Artificial Analysis Intelligence Index.
- [12]
MiMo-V2.6-Pro generates output at 129.7 tokens per second based on Xiaomi's API, well above the median of 77.4 t/s for open weight models of similar size.
- [13]
MiMo-V2.6-Pro has a time to first token of 2.17s based on Xiaomi's API, better than the 2.30s median for open weight models of similar size.
- [14]
Artificial Analysis compares open weights models only with other open weights models of the same size class, with Large defined as more than 150B parameters.
- [15]
The release coincides with the unveiling of the Alibaba V900 chip, which promises 500,000-chip clusters and is slated for production in Q1 2027.
- [16]
US export controls on Nvidia hardware ranged from restriction of H100 and A100 chips to the blocking of H200 imports in January 2026 and the closure of third-country cloud loopholes in May 2026.
- [17]
Chinese GPU and AI chipmakers captured 41% of the local accelerator market in 2025, up from near-zero in 2022, while Nvidia's China data center revenue has effectively dropped to zero.
- [18]
Builders can access the model through the Xiaomi MiMo open platform API, OpenRouter, or AI Studio.
- [19]
42 billion active parameters out of 1.02 trillion total is about 4.1% of the model active per token.
- [20]
At Xiaomi's listed prices, the 140M output tokens generated during the Intelligence Index run account for $121.80 of the $206.66 total, leaving $84.86, which corresponds to roughly 197 million input tokens if list prices applied with no cache-hit discount.
- [21]
Xiaomi's MiMo-V2.6 series was released on September 22, 2026, according to Forkast.
- [22]
Artificial Analysis lists MiMo-V2.6-Pro's release date as September 21, 2026.
- [23]
Artificial Analysis's FAQ gives MiMo-V2.6-Pro pricing on Xiaomi's API as $0.43 per 1M input tokens (better than average, median: $0.45) and $0.87 per 1M output tokens (very competitive, median: $1.68).
- [24]
The same Artificial Analysis page's model summary describes $0.43 per 1M input tokens as "somewhat expensive, median: $0.30" and $0.87 per 1M output tokens as "moderately priced, median: $1.13".
- [25]
When evaluated on the Intelligence Index, MiMo-V2.6-Pro generated 140M output tokens, which Artificial Analysis calls "somewhat verbose" in comparison to the median of 140M for open weight models of similar size.
- [26]
MiMo-V2.6-Pro's $0.87 output price is about 48% below the $1.68 size-class median cited in Artificial Analysis's FAQ.
- [27]
By pushing MiMo-V2.6-Pro to a score of 46 on the Artificial Analysis Intelligence Index v4.3, Xiaomi has effectively tied Grok 4.7, establishing a new ceiling for open-weight models.
- [28]
On the DeepSWE v1.1 benchmark MiMo-V2.6-Pro achieved 71.9%, trailing Claude Opus 5 at 74.0% but outperforming Claude Fable 5.
- [29]
MiMo-V2.6-Pro's performance on Terminal Bench 2.1 reached 89.9%, surpassing both Claude Opus 5 and GPT-5.6 Sol.
- [30]
The model recorded 94.0% on CyberGym, leading all reported models.
- [31]
The model achieved 53.1% on AutomationBench v1.0.6.
- [32]
Forkast states that these figures represent vendor-reported benchmarks and should be viewed with the standard caution regarding internal testing environments.
- [33]
Forkast's prior coverage reported that the training dashboard for this model showed a $432,000-per-day burn rate.
- [34]
Forkast argues that instead of stifling Chinese AI progress, the export control policies have acted as a catalyst for domestic independence.
Sources
2 independent publishers whose own reporting we read for this story.
- artificialanalysis.aiMiMo-V2.6-Pro - Intelligence, Performance & Price Analysis | Artificial Analysis
1 article · September 21, 2026
- forkast.newsXiaomi’s MiMo-V2.6 Ships Open Weights at Frontier-Class Performance – and the Timing Is Not an Accident – Forkast
1 article · September 22, 2026
Topics and entities
Topics
- LLM API Pricing
- Open-Weight Model Publishing
- Mixture of Experts Architectures
- US AI chip export controls