Xiaomi releases MiMo-V2.6 and open-sources a 1T-class flagship
Xiaomi finished the training run it put on a livestream: MiMo-V2.6 is now three shipped models, and two of them came out with their weights released the same day.
Xiaomi officially released the MiMo-V2.6 series on Monday and open-sourced the weights of MiMo-V2.6-Pro and MiMo-V2.6-Flash, alongside the technical report, the RL training environment, the training code, and a 9B distillation built on Qwen. Pro is the flagship — native multimodal, a 1M-token context window, and listed at more than a trillion parameters — and it holds its V2.5 API price at $0.435 per million input tokens and $0.87 per million output. Flash is a mixture-of-experts model with 309 billion total parameters that activates 15 billion per token, at $0.14 per million input tokens. Both also ship an "UltraSpeed" mode that Xiaomi says speeds up output by up to 20x, running the same checkpoint rather than a smaller one.
The framing is the part worth reading twice. Xiaomi places this release on what it calls the RSI — recursive self-improvement — path, and describes the work as scaling reinforcement learning on verifiable, complex tasks until the model expands its own boundary. The engineering disclosures match that ambition: the team froze the MoE router during training to suppress expert-load drift, and built a reward-hacking defence that spans reward design, adversarial evaluation, anomaly detection and verifier cross-checking. The multi-agent plumbing got rebuilt too — a unified trajectory representation with penalties, high-concurrency multi-agent interaction, and a data plane decoupled from the control plane so millions of trajectories can be moved without reworking the trainer.
The capability claims fan out well past chat. Xiaomi shows MiMo-V2.6-Pro decomposing an image or text prompt into a playable 3D scene with multiple agents programming, rendering and correcting each other; generating Blender assets; driving a Franka Panda arm from camera frames in closed loop; and running office software through native computer use. Two claims stand out for a lab doing this in the open: a materials-design case where the model screened metal-organic frameworks for PFAS capture before any specialist post-training, and a formal-proof case that ended with more than 6,000 lines of Lean verified by the Lean kernel with no unproven placeholders. Xiaomi also says Pro reaches parity with Claude Opus 5 and GPT-5.6 Sol on the Design Arena leaderboard.
Here is the take: the notable fact is not the benchmark sheet, it is the release policy. A trillion-parameter-class flagship with open weights, an open RL recipe and unchanged prices narrows the argument that frontier capability must stay behind an API — and Xiaomi is willing to name the self-improvement framing out loud while it does it. This is the same run we covered when the dashboard went live — Xiaomi is training MiMo-V2.6 in public — except now you can download the result. Since the model takes audio and video natively, our explainer on what multimodal AI actually means is the background.
Tim Dettmers' CMU lab has announced an open-source week built around the claim that frontier AI no longer needs a frontier budget — its inference framework runs a 125-billion-parameter Qwen 3.8 Flash Next on a single 24 GB GPU and DeepSeek V4.1's 550 billion parameters on a 128 GB MacBook, an AMD Strix box or an Nvidia DGX Spark. The concrete numbers: quantized inference of a Qwen 3.6 35B model at 450 tokens per second at 1.5 bits per weight, from kernels an agent optimized unattended on both CUDA and Metal. He also describes CliffCompaction, an auto-compaction method that runs sessions past a hundred million tokens, cuts cost roughly in half, and took one partner's total AI budget down 45%, plus a local research system he says beats Sakana AI's and Google's ScientistOne while offline.
The release itself is the thing to hold him to. Dettmers says the week's two open-source projects and four papers slipped by a day to ship as one package, so the artifacts are not out yet — the claims are his, on his blog, and they get tested when the repositories land. What is already clear is the argument underneath: with agents, the hard part of research moved from running experiments to maintaining a coherent ecosystem, and that is a problem cheap enough for a lab with a couple of GPUs.
What to watch: whether Xiaomi's Pro weights actually appear on Hugging Face at the listed scale, and whether dlab's published numbers survive the community's rerun.
Would you rather run a trillion-parameter flagship locally with no API bill, or trust a hosted model you cannot inspect? Tell us in the comments.
Sources: Xiaomi MiMo — MiMo-V2.6 release · Xiaomi MiMo · OpenRouter — MiMo-V2.6-Pro · Hugging Face — XiaomiMiMo · Hacker News discussion · TechNode · dlab Open Source Week — Tim Dettmers · Hacker News discussion
