Xiaomi Releases MiMo-V2.6 Open-Weight AI Models with Full RL Training Stack

Xiaomi has released two open-weight artificial intelligence models, MiMo-V2.6 Pro and MiMo-V2.6 Flash, under an MIT license, making weights available on Hugging Face, as reported by The Next Web (https://thenextweb.com/news/xiaomi-mimo-v2-6-open-weight-model-anthropic-distillation). Benchmark evaluation firm Artificial Analysis scored the flagship MiMo-V2.6-Pro at 46 on its Intelligence Index, placing it first among 114 open-weight models in its category and rating it the most cost-effective tracked system at $0.13 per task.

MiMo-V2.6-Pro uses a sparse mixture-of-experts architecture totaling 1.02 trillion parameters, with 42 billion active parameters per token to reduce operational overhead. The multimodal architecture supports text, image, audio, and video inputs with a 1 million-token context window. Xiaomi priced API access for Pro at $0.435 per million input tokens and $0.87 per million output tokens, alongside a high-throughput variant, Pro-UltraSpeed, that claims up to 20 times faster generation speeds at $4.35 input and $8.70 output per million tokens. For comparison, Anthropic cut Claude Opus 5.5 pricing to $4 per million input tokens and $20 per million output tokens on the same day.

The models were trained in a single continuous reinforcement learning run streamed live over less than six days. Pro and Flash completed 30 training steps across roughly 750,000 trajectories, with direct training compute costs totaling about $2.62 million for Pro and $0.85 million for Flash. Rather than isolating domains, Xiaomi unified software engineering, general agents, visual tasks, and cybersecurity workloads into a single training job. To counter reward hacking—where early agents bypassed bug-fixing by extracting solutions from future package updates—Xiaomi stripped build caches and forward Git history from training sandboxes and used a dedicated agent to detect remaining loopholes.

In Xiaomi's reported testing, Pro climbed from 58.4 to 72.57 on the DeepSWE benchmark across the training run, surpassing open-weight competitors including Moonshot's Kimi K3 and Alibaba's Qwen3.8 Max. Proprietary frontier models retain a lead: Artificial Analysis rates Claude Opus 5.5 at 58 compared to MiMo's 46, while on Terminal Bench 4.0, Pro scored 34.9 against Opus 5's 49.0 and OpenAI's GPT-6 Astra at 59.6. Xiaomi claims Pro matches Opus 5 and GPT-5.6 Sol across internal agent evaluations.

The smaller Flash variant is designed for high-throughput deployment, priced at $0.14 per million input tokens and $0.28 per million output tokens while retaining the 1 million-token context window and multimodal ingestion. In Xiaomi's benchmark reporting, Flash scored within four points of Pro on most internal agent tests, including 67.9 against 71.9 on DeepSWE, 52.3 against 53.1 on AutomationBench, and 61.2 against 62.0 on JobBench, while narrowly leading Pro on the CyberGym test with 95.1 to 94.0.

Alongside the weights, Xiaomi published its technical report, reinforcement learning framework, hyperparameters, data mixtures, costs, and over 7,000 task environments equipped with automated graders covering coding, security, administrative workflows, and music composition. Both models are available via Xiaomi's AI Studio, desktop and code tools, native API, OpenRouter, and Hugging Face. The release follows a September 10 threat report from Anthropic naming Xiaomi among seven Chinese organizations alleged to have distilled Claude outputs through automated requests, an allegation Xiaomi has not addressed.