Xiaomi MiMo-V2.6 Pro Tops Open-Weights: 1M Context and 20x UltraSpeed at Record Price-Performance

Xiaomi is back from a quiet corner with a loud claim. MiMo-V2.6 openly embraces RSI (recursive self-improvement — the model improving on verifiable tasks it creates for itself, like a chess player getting stronger by playing against itself) as its north star. The idea is not new, but Xiaomi pairs it with scaled reinforcement learning (RL) and tests it in open weights. Instead of leaning on one more pile of labeled data, the model advances through exploration and feedback on complex tasks where right and wrong can be checked. In practice, that means the lab grows the model with harder, measurable problems rather than just more data. For me, the series matters less as a point on a leaderboard and more as the first serious open rehearsal of self-improving training.

The lineup is not one model but three. At the top sits MiMo-V2.6 Pro — Xiaomi's most capable so far. Alongside it is Flash, tuned for speed, efficiency and cost. The third piece is Pro UltraSpeed mode, claiming up to 20x faster output at the same quality — a speedup that usually trades quality elsewhere, which makes the same-quality claim worth noting. On architecture, Flash is disclosed on GitHub as a Mixture-of-Experts (MoE — only the relevant experts fire per token, like calling the right specialist in a hospital rather than the whole staff) with 309 billion total parameters and 15 billion active per token. Pro keeps a higher active count, but the philosophy is shared: a small active network per token, low latency and cost. If you ask to summarize a 10-page PDF, the model wakes the summary and language experts, not the whole brain.

Price is why this story leads. MiMo-V2.6 Pro holds at 0.43 dollars per million input tokens and 0.87 per million output; Flash sits at 0.14 and 0.28. Xiaomi's claim is 1/120 to 1/160 of comparable international frontier pricing — the same job for roughly a hundredth of the cost. What does a 1-million-token window mean? About 700 to 800 pages in one go — like putting an entire season's scripts on the table and saying find the bug. That lets long legal contracts, large codebases or multi-step agent runs stay in one context without chunking. For a small team the takeaway is concrete: try Flash for free on OpenRouter (the video stresses the free access), switch to Pro when traffic grows; Flash is enough for hobby prototypes, Pro's window and stability shine in production.

Benchmarks tell a clear but qualified story. On Artificial Analysis' Intelligence Index, MiMo-V2.6 Pro scores 46, topping strong open-weights peers like Kimi K3 and Qwen 3.8 Max to claim the top open spot at release, a picture echoed by VentureBeat and Unite.AI under heads like better than DeepSeek. The channel's own World of AI benchmark confirms the same table and notes the model closing in on GPT-5.6 (referred to as GBT 6 Astra among closed frontiers). The map is not uniform, though: on DeepSway 1.1 and some visual-cyber niches, MiMo still trails closed leaders a touch. How to read it? State-of-the-art in general reasoning and coding, with room for fine-tuning in specialized perception and cyber cases. The 46 is the open-weights high-water mark; the real test is how it behaves off the leaderboard.

The skill set is what separates MiMo-V2.6 from a chat-only model. Xiaomi highlights three pillars: 3D spatial reasoning, multimodal perception and computer use. Turning text, an image or a video into a playable 3D world sounds ambitious; the demos bring theory closer to practice. A simplified mechanism is: 1) translate the prompt or reference visual into a scene graph, 2) materialize the scene as a Three.js app and a Blender scene, 3) connect that world to robot simulation or autonomous desktop tool use. Like an architect moving from sketch to maquette to construction robot. That chain powers product visualization, rapid game prototyping or factory simulation with one model and many outputs. The limit starts here too: scene generation impresses, but physics and interaction still want curation.

From Scene to Product: What the Live Demos Show

My favorite moment is the Sonic-style 3D game built from scratch by MiMo-V2.6 Pro. The model generates the 3D environment, character movement and physics in one pass and ties them into a cohesive loop with level design, obstacle placement and a camera system. The result is playable across multiple levels; tiny camera hitches aside, the loop holds. That compresses what would normally take a game team weeks or months of core prototyping into days of skeleton work. Say the acceleration curve feels too steep — the model keeps the same scene graph and tweaks the parameter, sculpting the existing world rather than starting over. My read is that MiMo is among the fastest open models at turning an idea into a playable draft in 3D.

On design, MiMo earns the not-a-frontier-clone label. Beyond Xiaomi's own blog examples, the creator prompts an Nvidia landing page. The 3D GPU model — a detail most models fumble — lands convincingly, accompanied by scroll triggers and interactive sections. There is even an overhead cutaway animating the cooling airflow, showing intake and exhaust paths. That is the difference a model with its own design taste makes in marketing and product pages: not a template feel, but a consistent language in lighting and typography. If most models wear the same off-the-rack cut, MiMo feels tailored; it does not clone every branded asset, yet it levels up in light and depth.

Simulation muscle peaks in the Formula drift scene. Done in Three.js as a single-page app, an F1-style car draws endless donuts; smoke, tire marks, multiple camera angles, reverse gear and slow motion are orchestrated inside one scene. The narrator calls it the best generation of its kind he has seen, closed Astra models included. The same strength shows in an isometric room: hover highlights animate the object and contrast with the environment. On the SVG side, a painting gains a living picture effect with trees swaying, wind and birds. The mechanism is simple and effective: bind scene, layers and particle systems to a single time-ordered track driven by scroll or hover. In practice, you can move from promo video to interactive storefront in one prompt.

Testing at art and city scale adds nuance. In art canvas and New York skyline generations, the model nails small details like tree lines, walkways, a helicopter and the pale halo of moonlight; even headlight beams are placed correctly. The snag appears on the city bridge: the deck overlaps buildings and the overpass connection slips. The flying cars you see are a layering bug, not a feature; depth ordering in dense urban scenes remains hard. That marks the gap between almost ready and shippable — strong overall composition, but architectural accuracy still wants a human pass. Even so, delivering this many layers coherently in one prompt is a breakthrough level for open weights.

Nostalgia, Calculation and Simulation: Detail in Every Pixel

A dose of nostalgia closes the circle with a demo many channels now skip as contaminated: a Windows 95 clone. MiMo codes floppy and desktop icons, a tips panel, a browser window and a DOS prompt in one generation, with a completeness that even some closed frontier models miss. Next comes a graphing calculator simulation and a sim game where parks, agriculture and housing scale as you ask to expand; mechanics and functions join the world as you grow it. This sums up why MiMo stands out for visualization and explanatory builds: you turn an abstract idea into an interactive maquette and keep adding floors. For a teacher, explaining fractals becomes a playable experiment rather than a static diagram.

Stepping back, MiMo-V2.6 Pro looks like a small version bump with outsized impact. Price stays flat while speed and quality rise together, the 1-million window holds, and the best open-weights label is secured. If Xiaomi's RSI and scaled RL scaffold truly works, a 3.0 mentioned alongside closed flagships would not surprise; if not, this remains an efficient intermediate step. Who should try it? Teams hunting rapid prototypes, 3D previews and cost-efficient production can start with Flash for free on OpenRouter; the Mimo chat front end and API expose the same core. Who should stay cautious? Institutions needing high regulation or strict data residency may love open-weights flexibility but should not ship without their own deployment and audit layer. My call is clear: worth trying, but do not let the allure of speed skip the verification layer in production.