Xiaomi is best known for making budget smartphones that undercut their competition on price without embarrassing themselves on specs. Now the company’s AI division has applied the same logic to large language models.
MiMo-V2.6-Pro launched on September 21 as a 1.02-trillion-parameter sparse mixture-of-experts model with 42 billion parameters active per forward pass — an omnimodal system covering text, images, video, and computer use. Its smaller sibling, MiMo-V2.6-Flash, runs 309B total parameters with 15B active. Both are MIT-licensed and open-weight.
The benchmark numbers are not politely competitive. MiMo-V2.6-Pro scores 46.32 on Artificial Analysis’s Intelligence Index, ranking first among open-weight models in its comparison class and ahead of Kimi K3 and Qwen3.8 Max — the models that have dominated the open-source coding and reasoning leaderboards for most of 2026. On agent benchmarks it reportedly runs even with Claude Opus 5 and GPT-5.6 Sol, which is a striking claim for a model from a company whose main business is selling $300 phones.
The detail worth pausing on is how much RL post-training moved the needle. On DeepSWE v1.1, Flash jumped from 48.8 to 65.68 and Pro from 58.4 to 72.57 — a 20–25% relative improvement from the same base architecture. The base model was already strong, but the RL stage is doing real work, not just polish. That pattern keeps appearing across the labs releasing competitive open models: the base compute is table stakes, and the training recipe after that is where differentiation actually happens.
MiMo-V2.6-Pro also introduces what Xiaomi calls “Vibe World” — a multi-agent capability for constructing and visually testing interactive 3D environments from image, video, or text prompts, including robotics control and music composition. The name is a bit much, but the underlying capability (coordinating specialized agents inside a shared visual simulation) is the direction the whole industry is heading toward for robotics and embodied AI research.
There is a larger pattern in this release. The gap between what a well-resourced tech company can produce in open weights and what the frontier labs offer as their flagship closed API has been narrowing all year. DeepSeek-V4.1-Flash (which arrived September 10), Kimi K3, and now MiMo-V2.6 all demonstrate that you don’t need to be OpenAI or Anthropic to produce frontier-competitive models — you need serious infrastructure, serious training talent, and apparently, sufficient motivation to go MIT-licensed rather than keeping it proprietary.
For practitioners, the practical question is whether 42 billion active parameters running on your own hardware at MIT terms competes with paying API rates for a closed model that’s only marginally better on the tasks you actually care about. For most real coding workloads, the answer is increasingly yes.