Author | Bi Weihao, Editor | Li Shuiqing
Zhidongzi reported on September 22 that this morning, Xiaomi's large model team released and open-sourced the new generation model Xiaomi MiMo-V2.6 series, which includes two native omni-modal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash. Xiaomi will also gradually open up MiMo-V2.6-Pro-UltraSpeed, which, compared to MiMo-V2.6-Pro, increases output speed by 20 times at the same intelligence level.
▲Official announcement of the Xiaomi MiMo-V2.6 series models (Source: X)
In addition, Xiaomi simultaneously open-sourced the RL environment and framework used for training the new models, as well as the model technical report.
Fuli Luo posted this morning stating that MiMo-V2.6 bets on large-scale RL scaling, conducting the largest single RL training run for open-source models to date. The team views this as a new step in exploring RSI (Recursive Self-Improvement) and paving the way to AGI. She stated frankly that the innovation and engineering difficulty of MiMo-V2.6 have already surpassed those of DeepSeek R1.
▲Fuli Luo's post (Source: X)
MiMo-V2.6-Pro scored 46.32 on the Artificial Analysis Intelligence Index, surpassing Kimi K3 and Qwen3.8 Max to become the highest-scoring open-source model on this leaderboard to date. However, in terms of scores, the overall gap with top open-source models is not large, and its score is comparable to the recently released Grok 4.7.
In the Code Arena WebDev leaderboard, which primarily tests AI web development capabilities, MiMo-V2.6-Pro scored 1628 in its first test, a significant increase of 153 points from the previous generation MiMo-V2.5-Pro's 1475 points. It currently ranks in the global top ten and third among open-source weight models.
In terms of pricing, the MiMo-V2.6 series maintains the API pricing of the V2.5 series. For MiMo-V2.6-Pro, the input (cache hit) is CNY 0.025 per million tokens, cache miss is CNY 3 per million tokens, and output is CNY 6 per million tokens; for MiMo-V2.6-Flash, the input (cache hit) is CNY 0.02 per million tokens, cache miss is CNY 1 per million tokens, and output is CNY 2 per million tokens; the pricing for MiMo-V2.6-Pro-UltraSpeed is ten times that of MiMo-V2.6-Pro.
A netizen shared a price comparison between MiMo-V2.6-Pro and Grok 4.7, showing that at the same intelligence level, the price of Grok 4.7 is approximately 29 times that of MiMo-V2.6-Pro.
▲Netizen comments (Source: X)
The warm-up for the release of the Xiaomi MiMo-V2.6 series was also very interesting. In the early hours of September 17, Fuli Luo posted real-time dashboards of the training data for the two models on social media, "live streaming" the training details of Xiaomi MiMo-V2.6 to the entire internet. The two rounds of training burned through an average of USD 30,800 per hour.
▲Fuli Luo's post live-streaming the training process (Source: X)
A netizen commented that this is one of the coolest open-source releases in history, not only live-streaming the training process but also open-sourcing the technical details, offering super strong performance while the pricing is absurdly low.
▲Netizen comments (Source: X)
However, some developers have also questioned the model's capabilities, believing that the MiMo-V2.6 series, at least the Flash model, has not met expectations. In OpenCode, it continuously loops executing commands and reading files, showing poor execution capabilities.
▲Netizen comments (Source: X)
Zhidongzi immediately conducted hands-on testing of the MiMo-V2.6 series models in OpenCode. From the actual results, the multimodal capabilities and programming proficiency of the MiMo-V2.6 series models are acceptable, but there are issues such as long thinking times and searching for directory files.
01.6 Days, 30 Steps, 750,000 Trajectories: From Coding to 3D and Scientific Research, Continuously Exploring RL Scaling
The large-scale RL scaling mentioned by Fuli Luo is presented as a series of numbers in the technical report of MiMo-V2.6. During its 6-day live-streamed training process, MiMo-V2.6-Flash and MiMo-V2.6-Pro each completed 30 steps, covering a cumulative total of approximately 750,000 trajectories, with corresponding training costs of about USD 850,000 and USD 2.62 million. After this round of training, the average pass rates for the training tasks of the two models increased by 25% and 12%, respectively. In the DeepSWE v1.1 test, Flash improved from 48.8 to 65.68, and Pro improved from 58.4 to 72.57, increasing by about 17 and 14 points, respectively.
During the RL training phase, MiMo-V2.6 invested a massive amount of computing power. After large-scale, multi-task reinforcement learning training, MiMo-V2.6-Flash and MiMo-V2.6-Pro achieved scores close to those of Claude Opus5 and GPT-5.6 Sol on most Agent Benchmarks.
In terms of the training system, the scaling of RL computing power is mainly reflected in three aspects: training scale, task environment, and infrastructure. Xiaomi adopted a fully asynchronous architecture, updating 1,568 samples in a single pass, supporting up to 1 million token contexts for training, and consuming approximately 3.5 billion to 3.7 billion tokens per RL step.
The training tasks are no longer limited to programming but cover scenarios such as general agents, vision, and cyber, mixing multiple harnesses to allow training signals generated from different tasks to jointly act on the model.
The reward mechanism has also been innovated. For long-horizon RL tasks, Xiaomi adopted an intra-group scoring mechanism, comparing the completion status and quality of different trajectories under the same task, and then reallocating reward signals. This can not only distinguish the difference between "completing the task" and "completing the task with high quality" but also guide the model to reduce the paths and tokens required to complete a single task.
As the training scale expanded, stability issues also emerged. Xiaomi froze the MoE Router to reduce model drift during training; for reward hacking, multi-layered defenses were set up from reward design and adversarial evaluation to validator cross-validation. At the engineering level, Xiaomi also unified trajectory representation and penalty mechanisms, and conducted high-concurrency interactions for different Agent frameworks. After the expansion of the training scale, the capability coverage of MiMo-V2.6 also expanded accordingly. The tasks handled by the new generation of models have extended further from programming to scenarios such as 3D content, computer operation, design, video, and music. For example, in a 3D game development task, the model can receive a video, an image, or a prompt, decompose the task independently, coordinate multiple agents to build 3D scenes and implement interactive logic, and conduct visual inspection and iteration based on the rendering results.
In Blender, it can also generate 3D assets based on text descriptions or reference images, and further use them for animation, 3D printing, and game development.
Through Computer Use, MiMo-V2.6 can enter embodied simulation environments. The model can receive multi-view camera feeds, continuously perform reasoning and decision-making, and then control the Franka Panda robotic arm through visual feedback to complete object grasping, color matching, and precise placement.
Design tasks are also within its capabilities. MiMo-V2.6 can generate complete front-end interfaces or presentations, including layout structures, components, interactive elements, and animations, and can collaborate with Figma as well as image and video generation tools.
The MiMo-V2.6 series models can also produce videos. In a product promotional video demonstrated by Xiaomi, everything from visual design, shot and animation arrangement, to music creation and beat-matching editing was completed by the model.
Music is a newly added capability. Xiaomi stated that MiMo-V2.6 has enhanced music understanding, aesthetic judgment, and music theory application, and has explored composition tasks for the first time.
The scientific research cases demonstrated by Xiaomi further showcase the transfer of general capabilities. Xiaomi explicitly stated that these scientific research tasks did not undergo specialized scientific research RL training but relied more on the model's existing general capabilities. In materials science, Xiaomi tasked MiMo-V2.6-Pro with designing a novel metal-organic framework (MOF) for adsorbing per- and polyfluoroalkyl substances (PFAS). Guided by materials experts, the model completed web searches, literature and patent reviews, hypothesis generation, and novelty assessment. Subsequently, it called open-source computational tools to build a simulation environment, calculated the binding strength between the MOF and PFAS, and finally screened out three candidate frameworks, which were then handed over to wet experiments for further verification.
In the direction of mathematics, it formalized the main theorem from Li and Yorke's classic paper "Period Three Implies Chaos" in Lean 4. After researchers designed exploration strategies, MiMo-V2.6-Pro advanced theorem statement and proof through sub-agent collaboration. The final project contains over 6,000 lines of Lean code and completed full verification through the Lean kernel, leaving no unfinished proof placeholders.
02.Racing Game Development and Xiaomi Auto Official Website Design: Prone to Long Thinking and Directory Browsing
Zhidongzi first tested basic multimodal capabilities by inputting a screenshot of a Hermes Agent application and asking MiMo-V2.6-Pro to recognize it without relying on tools. The model quickly completed the image reading task with good accuracy, but the speed was slightly slow.
With the same prompt, we changed the model and the screenshot, and asked MiMo-V2.6-Flash to execute this task. It can be seen that MiMo-V2.6-Flash took only 6 seconds, but the returned results were much less rich in content compared to the Pro model.
First, we used a web development task to test the web search, file manipulation, and front-end design capabilities of MiMo-V2.6-Flash, asking MiMo-V2.6-Flash to search for information independently and develop a demo for the Xiaomi Auto official website.
In this task, although the prompt we provided had very high requirements for web development, the overall development time for MiMo-V2.6-Flash, as a Flash-level model, was still too long. However, the quality of the web development was good; the interaction fluency of the webpage and image embedding were both fine, such as the test drive booking function demonstrated below.
However, the fly in the ointment is that for the Xiaomi Auto official website developed by MiMo-V2.6-Flash, although the text parts were all performance parameters of Xiaomi Auto, the searched and embedded images were of Mercedes-Benz, BMW, and Audi.
Finally, Zhidongzi inputted a classic racing game development prompt to test the game development capabilities of MiMo-V2.6-Pro. The prompt is as follows:
MiMo-V2.6-Pro took 64 minutes to complete the game development task. The developed game had quite a few issues: the track itself was not clear, the path was not strictly restricted, and in the initial state, the racing car model was buried underneath. The areas done relatively well were the game control feel and the overall lighting and shadow effects.
In addition, while testing the game development capabilities of MiMo-V2.6-Pro, Zhidongzi also inputted a 3D world game development prompt. Without relying on professional game development tools, the construction of the 3D world ultimately delivered by MiMo-V2.6-Pro was quite good. The functions, falling rain, and character movements were relatively reasonable, with only the minor issue of the umbrella not being in the character's hand.
Overall, from the hands-on testing, the MiMo-V2.6 series models give us a similar feeling to the previous generation—focusing on cost-effectiveness. The basic capabilities of the models are good, but there are some issues with thinking time and the route of task execution, making the overall model slightly weaker in task delivery efficiency. Of course, the hands-on tests conducted by Zhidongzi were all based on the models provided on OpenCode, which cannot represent the actual effects in Xiaomi's official Agent and API, nor can they fully represent the model's capabilities.
03.Conclusion: Xiaomi's New Work on the RSI Path But Model Capabilities Still Require More Hands-on Verification
From this release, it can be seen that MiMo-V2.6 focuses on the further scaling of large-scale RL training: completing 30 RL steps in 6 days and covering approximately 750,000 trajectories, with training tasks extending from Coding to Agent, vision, 3D, Computer Use, and scientific research tasks. At the same time, Xiaomi open-sourced the RL training environment, code, and technical report together, making public details such as the infrastructure and reward mechanisms during the training process.
From benchmarks to actual tests, MiMo-V2.6 has demonstrated strong basic capabilities, but there are still obvious differences between different tasks. Especially in long-horizon tasks, the model's thinking time, tool calls, and execution paths will directly affect the final delivery efficiency, and there are also many issues in the performance during the hands-on testing process. Where the model can ultimately go, besides the leaderboard scores, the quality and efficiency of task completion in subsequent actual use are equally important. Xiaomi hopes to explore RSI by continuously expanding the scale of RL training. Whether this path can further enhance the model's autonomous task capabilities still needs to be verified by subsequent versions and more actual tasks.