Xiaomi did two unusual things on September 21, 2026. It released MiMo-V2.6 Pro and MiMo-V2.6 Flash – two natively omnimodal models, open-sourced under the MIT license – and it trained them live in public, with a dashboard showing costs, token throughput, and benchmark scores as the RL run progressed. No US frontier lab does this. I pieced together the launch announcement, the API docs, the training livestream reporting, and the third-party benchmark trackers so you get one honest picture: what Pro, Flash, and the UltraSpeed variant each are, what the numbers actually say, and what open weights at this scale mean for people running local AI.

This guide covers the full V2.6 family – Pro vs Flash vs Pro-UltraSpeed on specs, price, and benchmarks – plus the training story, the "Vibe World" demos, and the practical question of what you can actually run yourself.

Source note: specs and prices below come from Xiaomi's MiMo-V2.6 announcement, the MiMo API model docs, and gateway listings (Vercel AI Gateway, OpenRouter). Scores are a mix of Xiaomi-reported RL curves and third-party trackers (Artificial Analysis, LLM Stats, Benchable) – labelled each time, because the two kinds of numbers mean different things.

The short version

What each variant is

SpecMiMo-V2.6-ProMiMo-V2.6-FlashPro-UltraSpeed
Total parameters~1T (reported ~1.02T)309BSame as Pro
Active per tokenNot disclosed (MoE)15B (hybrid attention MoE)Same as Pro
Context / max output1,048,576 / 128–131K1,048,576 / 128–131KSame as Pro
Input modalitiesText, image, audio, video, PDFText, image, audio, video, PDFSame as Pro
OutputTextTextText, up to ~20x faster
Speed techStandard servingEfficient MoEFP4 lossless quant + DFlash parallel decoding + TileRT kernels
Input / output per 1M$0.435 / $0.87$0.14 / $0.2810x Pro
Cached input per 1M$0.0036$0.0028–
LicenseMIT (open weights)MIT (open weights)Hosted only

The positioning is deliberate: Pro pushes the capability ceiling (Xiaomi says on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks), Flash holds most of the intelligence at roughly a third of the price, and UltraSpeed trades money for latency – up to 1,000 tokens/s – for realtime coding agents. Note the cache-read prices: at ~1/50th (Flash) to ~1/120th (Pro) of the miss rate, agents that reuse context – which is nearly all of them – cost far less than sticker. That is the same trick that makes DeepSeek V4.1 Flash so cheap in practice.

Trained in public: the RL livestream

This is the part of the release I find most important, more than any single score. Led by Luo Fuli – the ex-DeepSeek scientist Lei Jun hired to run the MiMo team, who previously led DeepSeek-V2 work – Xiaomi spent roughly half a year studying one problem ("how far RL can scale"), then ran the V2.6 RL phase in the open: a live dashboard with training progress, token consumption, cost meters, step timings, dataset composition across 23 categories, and benchmark curves updating mid-run.

The run itself: roughly 30 RL steps each for Pro and Flash across ~750,000 trajectories in under six days, on coding, general-agent, visual, and cybersecurity tasks, with 1,568 samples per update and contexts up to 1M tokens. Xiaomi says it froze the router to limit drift and used adversarial evaluation, anomaly detection, and verifier cross-checks against reward hacking. Final RL costs reported at ~$2.62M for Pro and ~$0.85M for Flash (a mid-run September 17 snapshot showed ~$512K at 49.4B Flash tokens – the meter kept running after that). Mid-training, Pro had already jumped from the V2.5 baseline of 19% to 65.97% on DeepSWE v1.1, sitting in a tier with Grok 4.6 (67%), Kimi K3 (69%), and Fable 5 (70%) – and the final numbers went higher still (see below).

Why this matters beyond Xiaomi: closed-door RL is the industry norm, which means nobody can check whether a capability jump came from the algorithm, the data, or the eval harness. A public cost-and-curve ledger – $128K-a-day-scale spending visible to everyone – is a reproducibility experiment as much as a marketing stunt. If you train your own models, the published environments and RL code are arguably the most valuable artifact of this release.

Benchmarks: where Pro leads and where it doesn't

Xiaomi's headline: 46.32 on the Artificial Analysis Intelligence Index v4.3 – its claimed highest open-source score, ahead of Kimi K3 and Qwen3.8 Max. Vendor framing aside, the RL curves are striking: DeepSWE v1.1 Pro 58.4 → 72.57, Flash 48.8 → 65.68. Third-party trackers add texture – Flash's reported strengths cluster in agentic and security work:

BenchmarkMiMo-V2.6-FlashMiMo-V2.6-Pro
CyberGym95.1%–
Terminal-Bench 2.187.6%–
OSWorld-Verified (computer use)80.8%–
MiMo Cyber Bench77.2%–
Toolathlon-Verified73.6%–
DeepSWE v1.165.68% (RL final)72.57% (RL final)
MiMo Coding Bench–63.2%
ProgramBench–26.5%

The honest read: this is the second open-weights family in a month – after DeepSeek V4.1 Flash – that looks genuinely competitive with closed flagships on agent work (terminal, computer use, repo coding) rather than just chat. The gaps are where you'd expect for efficiency-tuned open models: closed-book reasoning and the newest harnesses still belong to Astra and Fable 5.1. And cross-vendor score comparisons are directional – different harnesses, different effort settings, different dates. LLM Stats' head-to-head composite currently has Flash and Pro closely matched (45.7 vs 40.6), with Pro winning the three directly shared benchmarks.

Beyond chat: "Vibe World," robots, and real science demos

Xiaomi is pushing V2.6 past conventional software work into what it calls "Vibe World": from an image, video, or text prompt, the model coordinates agents to build and visually test interactive 3D scenes – Blender assets, frontends and presentations, assembled video, even music as scores and MIDI. The wilder demo: controlling a Franka Panda robotic arm from camera feeds. Research demos included screening materials for capturing PFAS "forever chemicals" and helping formalize a Lean 4 theorem in 6,000+ lines of kernel-verified code.

Treat demos as demos – carefully chosen, possibly retried – but the direction is consistent with the benchmarks: this family is built for agents that do things (browse, comprehend, reason, act) across modalities, not for winning trivia. The 1M native omnimodal context is what makes hour-long multimodal agent runs architecturally possible.

Price, access, and how to try it

What this means if you run local AI

First the honest answer: you are not running 1T – or even 309B – at home. Flash at 309B total needs datacenter-class VRAM in full precision; Pro at ~1T is firmly in the 2,000-GPU-plus-storage deployment world. My 12 GB RX 6800M will not be downloading either one – see how much VRAM LLMs need and the local-LLM GPU guide for what actually fits.

What you can use: the MIT weights mean community FP8 originals, GGUF quants, and vLLM/SGLang support will mature fast – and the API is among the cheapest frontier-class inference available (Flash at $0.14/$0.28 undercuts even DeepSeek V4.1 Flash's peak rates on input). The practical setup for local-AI folks: keep private and offline work on your local 7B–35B models, route heavy multimodal agent jobs to the $0.14 API, and watch the published RL code – the training transparency here teaches more than the weights alone. Our lab benchmarks show exactly where the local side wins: fast, private, repeatable generation with zero per-token cost.

Bottom line

MiMo-V2.6 is the most significant open-weights release of the month: a frontier-competitive agent family (46.32 AA Index, DeepSWE 72.6%) at open-model prices, MIT-licensed, with its training receipts published live. Flash is the value default for agent builders; Pro is the ceiling-pusher; UltraSpeed is for when latency is the product. And the livestreamed RL run sets a precedent I hope other labs copy – because "trust our curve" is no substitute for watching the meter run.

Sources