Xiaomi Releases MiMo-V2.6 with Trillion-Parameter Models and 7,000+ RL Environments

Xiaomi has launched the MiMo-V2.6 series, including the Pro model with trillion-scale parameters and the Flash model, which emphasizes cost efficiency. Both models can process text, images, video, and audio, support a context length of up to 1 million tokens, and are primarily designed for long-duration programming, tool usage, computer operations, and multimodal agents. More noteworthy than the models themselves is their training process: Xiaomi applied approximately 750,000 mixed reinforcement learning trajectories to each model, integrating programming, general agent tasks, vision, and cybersecurity into a single training session. Instead of relying solely on pass/fail scoring, the models learned by comparing multiple successful pathways. Training took less than six days, with public costs for Flash and Pro estimated at approximately $850,000 and $2.62 million, respectively. Third-party evaluator Artificial Analysis ranked Pro as the highest-scoring open-weight model currently available; however, it does not universally surpass leading proprietary models, as noticeable gaps remain in certain terminal operations and cybersecurity testing. Xiaomi has also released the weights under an MIT license, a 9B distilled model, a technical report, over 7,000 RL environments, and a training framework—providing a rare, comprehensive dataset for external research into whether scaling reinforcement learning can consistently enhance agent performance.

Article author and source: Xiaomi MiMo

MiMo-V2.6-Pro employs a sparse mixture-of-experts architecture with a total of 1.02 trillion parameters, activating approximately 42 billion parameters per token; Flash has a total of 30.9 billion parameters, activating approximately 15 billion parameters per token. Both support a context length of up to 1 million tokens and accept text, images, video, and audio. They also feature a five-layer multi-token prediction module to accelerate generation.

The term "native multimodal" here primarily describes the model's input understanding capabilities. The fact that publicly released weights currently output primarily text does not mean the model can directly generate images, videos, or natural speech. Many of the short videos, presentations, music, and 3D scenes demonstrated by the official team were created by the model writing code, operating software, and invoking external generation tools.

Pro aims to reach the upper limits of capability, while Flash seeks to handle similar tasks at a lower inference cost. The API pricing per million tokens is as follows: Flash costs $0.14 for input and $0.28 for output; Pro costs $0.435 for input and $0.87 for output. The official also offers Pro UltraSpeed, claiming unchanged quality with output speeds up to 20 times faster, but its input and output prices rise accordingly to $4.35 and $8.70. This "same quality" claim lacks independent verification.

Although the weights are licensed under the MIT license, both primary models far exceed the scale that can be easily deployed on typical consumer-grade hardware. "Open weights" addresses licensing issues related to inspection, modification, and self-deployment, but does not imply that these models can be run cost-effectively on personal computers.

Previously, teams often trained programming models, vision agents, and cybersecurity models separately, then attempted to combine their capabilities. Xiaomi summarizes this approach as “You Only RL Once”: programming, general agent, vision manipulation, and cybersecurity tasks, along with multiple agent execution frameworks, are mixed into a single reinforcement learning batch.

Each training step includes 1,568 prompts, with each prompt generating 16 trajectories, equivalent to approximately 25,000 trajectories; each of the two models completes 30 steps, resulting in approximately 750,000 trajectories per model. Each step processes about 3.5 to 3.7 billion tokens, with the longest tasks utilizing up to 1 million tokens of context.

The entire process took less than six days. The training cost for Flash was approximately $850,000, and for Pro, approximately $2.62 million. The average pass rate for training tasks improved by approximately 25% and 12%, respectively. On the long-term software engineering test of DeepSWE v1.1, which was not used for training, Flash increased from 48.8 to 65.7, and Pro increased from 58.4 to 72.6.

Long-term agent training faces a challenge: unit tests alone typically only indicate whether a task has been completed, but cannot distinguish between two different paths that both pass the test. One agent might quickly identify and fix the error, while another might perform hundreds of irrelevant searches before accidentally arriving at the same result.

MiMo-V2.6 introduces intra-agent scoring. The system first compares multiple trajectories generated for the same task to establish scoring criteria for that task; during online training, it ranks already successful trajectories and allocates more learning signals to solutions with higher quality, shorter paths, and fewer token expenditures.

Xiaomi refers to this mechanism, in which comparison signals are generated from the model’s own samples, as a "self-improvement loop." However, this does not mean the model modifies itself autonomously after deployment without human involvement; the training environment, tasks, validators, reward structures, and computational resources are still provided by the research team, and weights are updated through formal training procedures.

As the training scale expanded, the team also froze the mixture-of-experts router to reduce drift between different training and inference environments. They further minimized opportunities for the Agent to exploit scoring loopholes through adversarial evaluation, anomaly detection, environment hardening, and cross-checking by multiple validators. After training, the MOPD² distillation method reuses teacher trajectories and historical prefixes from human demonstrations to focus training on critical decision points, eliminating the need to regenerate complete task histories each time.

Artificial Analysis awarded MiMo-V2.6-Pro a score of approximately 46, making it the highest-scoring open-weight model from the institution at that time. In official public tests, Pro achieved a score of 71.9 on DeepSWE v1.1, nearing the 74-point scores of DeepSeek V4.1 Flash and Claude Opus 5, and surpassing Claude Fable 5’s 70 points.

On AutomationBench, Pro scored 53.1, slightly above GPT-6 Astra’s 52.0, Claude Opus 5’s 50.3, and Claude Fable 5’s 46.2, but slightly below DeepSeek V4.1 Flash’s 54.8. GDPval-AA’s professional task score is 1673, close to Claude Opus 5’s 1708.

These results should not be interpreted as Pro having fully surpassed closed-source state-of-the-art models. On Terminal Bench 4.0, Pro scored only 34.9, significantly lower than GPT-6 Astra’s 59.6 and Claude Fable 5.1’s 55.1. In the ExploitBench cybersecurity test, Pro scored 47.9, while GPT-6 Astra, GPT-5.6 Sol, and Claude Fable 5 achieved scores of 100, 78.5, and 78.0, respectively.

Some visual and cybersecurity results are derived from Xiaomi's internal benchmarks, which cannot be fully replicated by external parties; Artificial Analysis's composite index primarily evaluates English text tasks and cannot demonstrate that the model's audio, video, and robotics capabilities are equally superior.

Xiaomi refers to the direction of extending model generation beyond software as "Vibe World." Agents can analyze requirements from text, images, or videos, enabling multiple sub-agents to collaboratively build 3D scenes, code interactive logic, and run rendering. They then refine the results based on visual feedback until delivering a fully functional interactive world.

The model can also operate Blender to generate assets usable for animation, gaming, or 3D printing. In robotic simulation, it receives multiple camera feeds and uses visual feedback to continuously control the Franka Panda robotic arm, completing tasks such as grasping, color matching, and precise placement.

These are officially selected demonstration cases, not large-scale robotic success rate tests. In particular, closed-loop control in simulation environments cannot directly prove that the model will maintain the same performance on real robots when facing friction, lighting changes, occlusions, and sensor errors.

In a materials research case, Xiaomi researchers asked Pro to identify novel metal-organic framework materials capable of adsorbing PFAS "forever chemicals." The model retrieved relevant papers and patents, formulated hypotheses, assessed novelty, and then invoked open-source computational tools to build a simulation environment to calculate the binding strength between candidate materials and PFAS, ultimately screening three structures worthy of advancing to wet-lab experiments.

This indicates that the model can chain together literature retrieval, environment setup, computational simulation, and result screening; however, the candidates are still computational results and have not yet been experimentally validated as new materials in the lab.

Another example is formalized mathematics. After researchers provided exploration strategies, Pro collaborated through multiple sub-agents to encode Li and Yorke’s classic result “period three implies chaos” into over 6,000 lines of Lean 4 code. The final proof was verified by the Lean kernel without any incomplete placeholders; however, the official team explicitly noted that researchers subsequently revised and integrated the results, making it a human-machine collaborative effort rather than a fully autonomous proof generated by the model from scratch.

This同步 release also includes MiMo-V2.6-Distill-Qwen-9B, a 9B-parameter agent model fine-tuned under supervision using data generated by MiMo, based on Qwen3.5-9B. It is designed for programming, general tools, visual programming, and cybersecurity, and serves as a compact starting point for RL research.

Xiaomi also announced the release of technical reports, over 7,000 reinforcement learning task environments, an end-to-end RL framework, and composable small-agent execution tools. This is more valuable for research than simply uploading final weights, as external parties can now determine whether performance improvements stem from the model, task distribution, validator, or agent framework.

However, these resources have just been released and have not yet undergone a systematic audit by the community regarding the integrity of the training data, environmental reproducibility, and security. Whether the official claim of “complete openness” is sufficient to reproduce identical results on other computing platforms still requires practical verification.