本文来源:https://www.sfccn.com/2026/9-23/zNMDE0MDdfMjI1MDAzNw.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Source of this article: https://www.sfccn.com/2026/9-23/zNMDE0MDdfMjI1MDAzNw.html For copyright issues, please notify us by email: 13951431913#139.com or submit a comment on the right.
Home > Business > Main Content: Xiaomi's Large Model Burns Through $3.47 Million in 6 Days 2026-09-23 10:57:22 21st Century Business Herald 21st Century Finance APP Lei Chen 21st Century Business Herald Correspondent Lei Chen On September 22, Xiaomi officially released and open-sourced the MiMo-V2.6 series, which includes the full-modality flagship Pro and the high-efficiency reasoning Flash. The Pro ultra-high-speed mode V2.6-Pro-Ultraspeed was also launched simultaneously, and the official version of the MiMoDesktop desktop client was released. Lei Jun, founder of Xiaomi Group, posted on Weibo the same day stating that since September 17, the post-training process of reinforcement learning for MiMo-V2.6 had been publicly livestreamed. The six-day livestream was backed by nearly half a year of foundational research and engineering trial-and-error. In this iteration, the model architecture and parameter scale were kept unchanged; by expanding the scale of RL training, capabilities in coding, agents, appearance, and research tasks were improved, while API pricing remained unchanged. According to the Comprehensive Intelligence Index released by Artificial Analysis, MiMo-V2.6-Pro scored 46 points, ranking first among global open-source models and first among domestic models, surpassing Kimi K3 and Qwen3.8 Max. It is the first to support full modality including text, images, video, and audio, making it the only full-modality Pro model in the global open-source lineup. The single-task cost leaderboard indicates that the actual measured single-task cost is $0.13, entering the $0.1 range for the first time. API pricing follows V2.5, with the selling price being only 1/20 to 1/60 that of overseas models at equivalent intelligence levels. The scale of expenditure is also intuitive: RL post-training took approximately 6 days, with Flash costing about $850,000 and Pro about $2.62 million, totaling over $3.47 million and accumulating approximately 750,000 trajectories. In long-horizon software engineering tests, Flash's score increased from 48.8 to 65.68, and Pro's score rose from 58.4 to 72.57. On the process side, Xiaomi utilized a large Batch and full asynchronous architecture, updating 1,568 samples per step, compatible with 1M context training, with single-step tokens reaching 3.5-3.7B. A hybrid multi-Harness framework was employed to provide more precise reward signals for Long Horizon RL; a frozen Router was used to suppress drift, and a Reward Hacking defense line covering reward appearance, adversarial testing, anomaly detection, and validator cross-checking was established. The capability leap is evident in scientific research scenarios. Acting as a "Co-Scientist," Pro assisted material teams in generating MOF ideas to adsorb PFAS, shortening the R&D cycle from about one month to 2–3 days, increasing efficiency by approximately 10 times. It also formalized the main theorem of the Li-Yorke theorem in the paper "Period Three Implies Chaos" in Lean 4, generating over 6,000 lines of source code, all verified by the kernel, and without prior Lean-specific post-training. In Design Arena actual tests, its appearance capabilities are comparable to Claude Opus 5 and GPT-5.6 Sol. The intensity of open-sourcing is another focus. In addition to releasing the Pro and Flash weights, Xiaomi has also open-sourced RL process results, MiMo-V2.6-Distill-Qwen-9B, over 7,000 high-quality task environments, and an end-to-end RL training framework along with a lightweight Agent framework. Starting RL training from the Distill-Qwen-9B baseline, all 11 measured metrics improved: SWE-bench Verified score rose from 61.1 to 66.2, MiMo Cyber Bench went from 31.3 to 47.0, Terminal Bench 2.1 increased from 37.1 to 52.8, and MiMo Visual Coding climbed from 64.0 to 72.4. Objectively, this is a combination of "heavy investment in the training end, low-price output in the inference end, and comprehensive release in the open-source end." The $3.47 million covers only RL post-training, excluding preliminary R&D and inference costs; whether the $0.13 single-task cost can support a commercial closed loop, and whether open-sourcing can translate into ecosystem dominance, remain to be seen. However, Xiaomi has at least placed "how much potential can be unlocked by post-training" on the livestream and open-source stage.