← Back to feed

Xiaomi released MiMo-V2.6-Pro on September 21, and it immediately landed atop Artificial Analysis’ Intelligence Index with a score of 46 — tied with xAI’s Grok 4.7, which debuted the same day, and above every other open-weight model currently tracked.

The model sits above Grok 4.6 (44), Gemini 3.8 Flash (41), DeepSeek V4.1 Flash (39), and DeepSeek V4.1 Pro (36) on the AA index. It also clears both Kimi K3 and Qwen3.8 Max, which had been the prior open-weight leaders. Artificial Analysis places it squarely on the intelligence-vs-cost Pareto frontier: no cheaper model offers more intelligence, and no lower-intelligence model offers better cost.

Architecture and Scale

MiMo-V2.6-Pro runs 1.02 trillion total parameters. The companion MiMo-V2.6-Flash comes in at 309 billion parameters, targeting high-volume production workloads at lower cost. Xiaomi is also shipping MiMo-V2.6-Pro-UltraSpeed, which it says reaches up to 20 times the normal Pro output speed for latency-sensitive use cases.

Both Pro and Flash are natively omnimodal: text, image, audio, and video on a 1-million-token context window. Coding, visual reasoning, and computer-use tasks were cited as primary design targets.

Training at Scale

Xiaomi trained V2.6 across 750,000 reinforcement-learning trajectories spanning programming, vision, and counterfactual tasks. To reduce training drift, the team froze the mixture-of-experts router during RL — a departure from standard practice. The company is open-sourcing more than 7,000 RL environments and training tools under the MIT license alongside the model weights.

Fuli Luo, a former DeepSeek researcher now heading the MiMo team, said V2.6 is likely one of the largest single reinforcement-learning runs for an open-weight model.

Pricing

On Xiaomi’s API:

• Pro: $0.435/M input, $0.87/M output
• Flash: $0.14/M input, $0.28/M output

Artificial Analysis measures the effective cost at $0.13 per Intelligence Index task for Pro, confirming its position in the upper-left quadrant of the price-performance chart. Output speed on standard Pro is approximately 134 tokens per second.

Weights are available on Hugging Face under the MIT license — free to download, fine-tune, and run on any hardware without per-token fees.

Context

Arena.ai added mimo-v2.6-pro to the Code Arena: WebDev leaderboard on September 21. Full text and coding Arena results are pending. The model’s appearance on the same day as Grok 4.7 — and with an identical AA score — makes a direct comparison the obvious next data point once human preference evals accumulate.

V2.6 is Xiaomi’s third major open-weight release in roughly a year. The V2.5-Pro scored 78.9% on SWE-bench Verified; Xiaomi has not yet published a V2.6 SWE-bench number, so the agentic capability picture remains incomplete.