MiMo-V2.6
A series of agentic reinforcement learning models released by Xiaomi, comprising Pro and Flash variants. The series focuses on software engineering agent tasks and employs the RLVR training paradigm.
Core Facts
Xiaomi officially disclosed that the total training cost of MiMo-V2.6 was only $3.5 million.
The live-stream page simultaneously trains two models, MiMo-V2.6 Pro and MiMo-V2.6 Flash.
Xiaomi has publicly released MiMo-V2.6's model weights, technical report, reinforcement learning (RL) environment, and training code.
Timeline (last 90 days)
MiMo-V2.6 follows a technical approach of having the model evolve through reinforcement learning.
Xiaomi officially stated that at an equivalent level of intelligence, MiMo-V2.6's cost is one-twentieth to one-sixtieth that of leading overseas models.
MiMo-V2.6 scored 34.9 on command-line task tests, below GPT-6 Astra's score of 59.6.
In an F1 racing drift demo, MiMo-V2.6 generated a single webpage file outputting over 43,000 tokens, with a generation cost of approximately 0.2 yuan.
MiMo-V2.6 performs real-time computation and control of a robotic arm placing blocks in NVIDIA's physics simulation environment using footage from two cameras on the wrist and the tabletop.
Xiaomi's MiMo series is an open-source reasoning model released by Xiaomi; MiMo-V2.5 demonstrates strong performance on code and mathematical reasoning tasks and employs reinforcement learning for post-training optimization
MiMo-V2.6 received day-0 support from vLLM on its release day, with both version sizes being deployable via vLLM on the day of release.
MiMo-V2.6 and V2.5 belong to the same architectural category, and previously deployed server-side components can be used directly, including hybrid sliding window attention + global attention, the MiMo reasoning and tool-call parser, and DFlash speculative decoding.
MiMo-V2.6 has a built-in DFlash speculative decoding mechanism that can draft 7 tokens per step.
MiMo-V2.6 uses native FP8 weights.
28 more timeline events
All Facts (20)
Xiaomi officially disclosed that the total training cost of MiMo-V2.6 was only $3.5 million.
80%VerifiedThe live-stream page simultaneously trains two models, MiMo-V2.6 Pro and MiMo-V2.6 Flash.
65%Partially VerifiedXiaomi has publicly released MiMo-V2.6's model weights, technical report, reinforcement learning (RL) environment, and training code.
75%UnverifiedMiMo V2.6 Pro uses a Mixture of Experts (MoE) architecture, and the Flash version is likewise a MoE architecture.
70%UnverifiedMiMo-V2.6 covers multiple modalities including text, image, and speech, making it an all-modality model.
70%UnverifiedIn an F1 racing drift demo, MiMo-V2.6 generated a single webpage file outputting over 43,000 tokens, with a generation cost of approximately 0.2 yuan.
50%UnverifiedMiMo-V2.6 scored 34.9 on command-line task tests, below GPT-6 Astra's score of 59.6.
50%UnverifiedXiaomi officially stated that at an equivalent level of intelligence, MiMo-V2.6's cost is one-twentieth to one-sixtieth that of leading overseas models.
50%UnverifiedMiMo-V2.6 follows a technical approach of having the model evolve through reinforcement learning.
50%UnverifiedMiMo-V2.6 performs real-time computation and control of a robotic arm placing blocks in NVIDIA's physics simulation environment using footage from two cameras on the wrist and the tabletop.
50%UnverifiedXiaomi's MiMo series is an open-source reasoning model released by Xiaomi; MiMo-V2.5 demonstrates strong performance on code and mathematical reasoning tasks and employs reinforcement learning for post-training optimization
50%UnverifiedMiMo-V2.6 uses native FP8 weights.
50%UnverifiedMiMo-V2.6 has a built-in DFlash speculative decoding mechanism that can draft 7 tokens per step.
50%UnverifiedMiMo-V2.6 received day-0 support from vLLM on its release day, with both version sizes being deployable via vLLM on the day of release.
50%UnverifiedMiMo-V2.6 and V2.5 belong to the same architectural category, and previously deployed server-side components can be used directly, including hybrid sliding window attention + global attention, the MiMo reasoning and tool-call parser, and DFlash speculative decoding.
50%UnverifiedThe MiMo-V2.6 model supports a context length of 1M (one million) tokens.
50%UnverifiedThe team achieved bitwise-matched consistency between rollout and the trainer during the validation stage, ensuring that model computations during inference and training are numerically identical.
50%Unverified310B parameters occupy approximately 620 GB of storage in float16 or bfloat16 precision; completing a full synchronization within 2 seconds implies that the effective bandwidth must reach 310 GB/s or more.
50%UnverifiedThe system uses vLLM to drive the rollout (sample generation) process.
50%UnverifiedThe transfer of all 310B parameters over the ICI (inter-chip interconnect) network completes within 2 seconds.
50%