Why Xiaomi’s MiMo-V2.6-Pro Now Leads Open-Weight AI Models

MiMo-V2.6-Pro’s debut matters because it combines a leading open-weight benchmark result with unusually low listed API prices and a substantial release of post-training artifacts. The more cautious reading is equally important: Xiaomi has reached the top of the open-weight category on one widely tracked composite benchmark, not the overall capability frontier.

The result: first among open weights, not first overall

Artificial Analysis lists MiMo-V2.6-Pro as its highest-ranked open-weight model with an Intelligence Index score of 46. The same ranking places GLM-5.3 at 45 and Kimi K3 at 44. 31

That score is also in the range reported for xAI’s proprietary Grok 4.7 configuration. But Xiaomi itself says its flagship still trails the strongest closed models. Artificial Analysis lists Claude Fable 5.1 and GPT-6 Astra at 53, a seven-point difference from MiMo’s 46. 119

This makes the milestone meaningful, but specific: Xiaomi has set a new open-weight reference point rather than demonstrated overall parity with the best closed systems.

Why MiMo-V2.6-Pro improved

Xiaomi calls its post-training approach “You Only RL Once.” Rather than training separate policies for coding, general agents, visual tasks, and cybersecurity, it mixes those domains—and multiple agent harnesses—into a single reinforcement-learning run. Xiaomi’s premise is that the same policy can learn transferable behaviors such as identifying missing information, choosing tools, assessing tool outputs, and recovering after a failed attempt. 5

The company says both Pro and Flash completed 30 reinforcement-learning steps over roughly 750,000 trajectories in under six days. Reported training costs were about $2.62 million for Pro and $850,000 for Flash. Those figures are Xiaomi’s own reported costs, not independently audited totals. 10

The result suggests that scaling reinforcement learning across diverse, tool-using environments can produce a large capability gain without building a bespoke post-training pipeline for every task category. Xiaomi’s claimed 20-point increase over MiMo-V2.5-Pro should still be treated as a release-era comparison rather than proof of general superiority across every real-world workload. 14

Cost is central to the story

MiMo-V2.6-Pro’s listed API price is $0.435 per million input tokens and $0.87 per million output tokens. A comparison published using Artificial Analysis’ task-cost estimates puts Pro at roughly $0.13 per Intelligence Index task, versus $3.74 for Grok 4.7 xHigh. 21

Benchmark task cost is not the same as the cost of a production workload: it depends on prompt length, output length, caching, tool usage, and inference settings. Still, the combination of a 46 score and low token prices is why Pro appears attractive for teams that need strong reasoning or agentic performance at scale.

MiMo-V2.6-Flash: the lower-cost sibling

MiMo-V2.6-Flash is Xiaomi’s efficiency-oriented model. It uses a sparse mixture-of-experts architecture with 309B total parameters and 15B active parameters, compared with Pro’s reported 1.02T total and 42B active parameters. Both models support a 1 million-token context window and accept text, image, video, and audio inputs. 4

Flash is priced for higher-frequency deployment at $0.14 per million uncached input tokens and $0.28 per million output tokens; Xiaomi lists cached input at $0.0028 per million tokens. 310

Xiaomi’s published benchmark comparisons put Flash below Pro on DeepSWE v1.1, while positioning it as a lower-cost option for long-context and multimodal workloads. These are vendor-published comparisons, so they are useful for understanding Xiaomi’s target positioning but should be validated against independent tests for a particular deployment. 10

An early practitioner view of Flash

Tim Dettmers described Flash as feeling like “the best model in the 300B to 550B class,” saying it performed better than DeepSeek V4.1 and GLM-5.3 Flash in his use. He also praised its automatic context compaction in a session exceeding 3 million tokens and reported about 250 tokens per second decode and 2.6k tokens per second prefill. 44

Those observations are notable, especially for long-running coding workflows, but they are one practitioner’s report—not an independent benchmark or a guarantee of performance on other hardware and serving stacks.

What Xiaomi is actually releasing

The release goes beyond downloadable weights. Xiaomi says it is open-sourcing Pro and Flash under an MIT license alongside a technical report, more than 7,000 RL task environments, an end-to-end RL framework, and composable mini-harnesses. It also exposes a dashboard sourced from trainer logs for the two RL runs. 462

That level of post-training visibility can help researchers inspect the training approach and build on the task environments. It does not, by itself, amount to a complete reproduction package for the full model-development process: the provided materials do not establish that every pretraining dataset, infrastructure component, or training detail has been released.

What this says about the open-weight race

MiMo-V2.6-Pro strengthens the evidence that Chinese labs are highly competitive—and currently leading on this particular open-weight Intelligence Index ranking. Its score exceeds the listed open-weight scores for GLM-5.3 and Kimi K3, while its availability under an MIT license makes it materially different from a closed API-only model. 3136

But it would be an overreach to turn that finding into a broad claim that Chinese labs lead U.S. labs overall. The same index places the top closed models, Claude Fable 5.1 and GPT-6 Astra, at 53. 19

The practical takeaway is simpler: open-weight AI has moved closer to expensive proprietary systems on this benchmark, and MiMo-V2.6-Pro is now a consequential option for teams that value deployable weights, low inference pricing, multimodal input, and long-context agent workflows.