MiMo V2.6 Pro
A large language model released by Xiaomi, adopting a sparse Mixture-of-Experts architecture with 472 billion total parameters and 31 billion activated parameters. It supports a context window of one million tokens and is positioned as an open-source, cost-efficient model.
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)
The open-source large model MiMo V2.6 Pro uses the MIT license.
The MiMo V2.6 Pro cluster runs a hello test at approximately 15 tokens/second, dropping to approximately 12 tokens/second for complex tasks.
In a knowledge quiz about Craig David's songs, the 3-bit quantized version gave more accurate answers than the full-parameter Pro version.
The MiMo V2.6 Pro cluster supports multimodality and can recognize Age of Empires II screenshots and build a menu system.
The 3-bit quantized version of MiMo V2.6 Pro produced a black screen for a solar system generation task and reported a runtime error for a nuclear explosion scene.
Manually finding the gray node where the model ends its thinking and clicking the end think tag can forcibly terminate the thinking loop.
The blogger recommends using no lower than 8-bit quantization for MiMo V2.6 Pro, even for the Flash version.
MiMo V2.6 Pro is quite sensitive to quantization — the more aggressively it is compressed, the more coding errors occur and the worse the experience becomes.
MiMo V2.6 Pro achieved a score of 71.9 on the SWE benchmark, surpassing GLM's score of 66.9.
MiMo V2.6 Pro requires approximately 573 GB of memory to run.
22 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 Flash has 310B total parameters and 15B activated parameters, with a speed nearly twice that of Pro.
70%UnverifiedMiMo V2.6 Pro uses a Mixture of Experts (MoE) architecture, and the Flash version is likewise a MoE architecture.
70%UnverifiedMiMo V2.6 Pro has 1.02T total parameters and 42B activated parameters.
70%UnverifiedThe MiMo V2.6 Pro cluster runs a hello test at approximately 15 tokens/second, dropping to approximately 12 tokens/second for complex tasks.
50%UnverifiedManually finding the gray node where the model ends its thinking and clicking the end think tag can forcibly terminate the thinking loop.
50%UnverifiedThe 3-bit quantized version of MiMo V2.6 Pro produced a black screen for a solar system generation task and reported a runtime error for a nuclear explosion scene.
50%UnverifiedThe MiMo V2.6 Pro cluster supports multimodality and can recognize Age of Empires II screenshots and build a menu system.
50%UnverifiedIn a knowledge quiz about Craig David's songs, the 3-bit quantized version gave more accurate answers than the full-parameter Pro version.
50%UnverifiedThe open-source large model MiMo V2.6 Pro uses the MIT license.
50%UnverifiedMiMo V2.6 Pro requires approximately 573 GB of memory to run.
50%UnverifiedMiMo V2.6 Pro achieved a score of 71.9 on the SWE benchmark, surpassing GLM's score of 66.9.
50%UnverifiedMiMo V2.6 Pro is quite sensitive to quantization — the more aggressively it is compressed, the more coding errors occur and the worse the experience becomes.
50%UnverifiedThe blogger recommends using no lower than 8-bit quantization for MiMo V2.6 Pro, even for the Flash version.
50%UnverifiedIn 3D origami reasoning tasks, Grok 4.7 is clearly superior to MiMo V2.6 Pro.
50%UnverifiedIn a 500-word first-person writing test, MiMo V2.6 Pro performed the best, surpassing both Grok 4.7 and Flash.
50%UnverifiedMiMo V2.6 Pro's total parameters are approximately half those of Kimi K3, but the activated parameters are on the same order of magnitude.
50%UnverifiedMiMo V2.6 Pro has a context window of 1 million tokens, on the same tier as Gemini 3.5 and Kimi K3.
50%