--- title: "Xiaomi open-sources MiMo-V2.6, a multimodal model that writes video, images, and audio" date: 2026-10-31 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-xiaomi-mimo-v26 language: en --- # Xiaomi open-sources MiMo-V2.6, a multimodal model that writes video, images, and audio > Xiaomi's MIT-licensed MiMo-V2.6 (1T-param MoE) generates video, images, and audio from text — and anyone can self-host the weights. ## Key takeaways - Xiaomi (小米) released MiMo-V2.6 in September 2026 under an MIT license with open weights. - The Pro variant is a ~1.02-trillion-parameter mixture-of-experts (42B active); Flash is 309B (15B active). - It is natively multimodal: text, image, video, and audio in one architecture, with a 1-million-token context. - API rates run about US$0.435/$0.87 per million tokens (Pro) and US$0.14/$0.28 (Flash). - For creators, the draw is end-to-end video and image generation they can run or fine-tune privately. ## One model, many media types Most generative-media (生成式媒体) stories are about a single trick: a video model, an image model, a voice model, each in its own walled garden. Xiaomi's MiMo-V2.6, released in September 2026, argues the opposite — that a single open model should handle text, image, video, and audio together, and that creators should be allowed to download it. The release spans a ~1.02-trillion-parameter mixture-of-experts Pro (42 billion active parameters) and a 309B Flash (15B active), all under an MIT license with weights published for commercial use and fine-tuning. The headline for a creative audience is not the parameter count but the workflow it implies. Because the model is natively multimodal, a user can drop in a screen recording and ask for a replicated web page, feed an image and get spoken commentary, or describe a short film and receive visual design, camera and motion planning, and a synchronized narration. Xiaomi demonstrates end-to-end video creation — from concept and storyboard to scored, voiced output — and image generation with consistent style and layout. Audio, including music scoring, is part of the same toolbox. The point is a continuous pipeline rather than four separate subscriptions. ## Why "open" is the actual feature MiMo-V2.6 ships more than weights. Xiaomi published the technical report, training environments, and reinforcement-learning code, plus over 7,000 RL task environments for agent work. For media creators, the practical upside is control: you can host the model on your own hardware, tune it on proprietary brand assets, and avoid sending unreleased material to a third-party cloud. In markets where data residency and IP protection matter, that is often the difference between experimentation and adoption. The economics are deliberately aggressive. API pricing sits at roughly US$0.435 input / US$0.87 output per million tokens for Pro, and US$0.14 / US$0.28 for Flash — among the lowest rates for a frontier-class multimodal model. Xiaomi also reports the final Pro training run cost about US$2.62 million over under six days, a figure meant to show that open training at scale is no longer the exclusive domain of the largest labs. On third-party benchmarks, the Pro variant reached the top of open-weight models on the Artificial Analysis intelligence index at launch, though vendor-adjacent numbers should be read as indicative. ## What this means for Chinese generative media MiMo-V2.6 sits inside a wider Chinese push to make foundation models (大模型) both capable and open. Zhipu's GLM-Image, Alibaba's Qwen family, and Tencent's Hunyuan line all publish weights or open APIs, and the competitive effect is a falling cost curve for every media task downstream. Where a year ago a studio might have licensed three separate tools, a team can now assemble a pipeline from open models — image from one, video from another, orchestrated by an open multimodal brain like MiMo. The risk, familiar to anyone who has shipped AI content, is coherence across the pipeline. A model that writes a script, draws a character, and scores music in separate steps can still drift: the face changes between shots, the voice loses a beat, the music fights the dialogue. Xiaomi's answer is architectural — cross-modal reasoning built in, not bolted on — but real-world consistency still depends on how creators wire the steps together. MiMo-V2.6 is a strong engine; it is not yet a finished editing suite. ## The honest constraints This is a general-purpose model, not a specialist media tool. It will not beat a focused video model like Kling or a focused image model like GLM-Image on their home turf, just as those will not match MiMo's reasoning and coding breadth. The "music" capability, in particular, should be read as demo-grade composition help rather than a replacement for a producer. And although the weights are open, running the trillion-parameter Pro variant at full quality still demands serious hardware; the distilled and Flash tiers exist precisely for that reason. ## Honest limitations This article draws on Xiaomi's release materials and third-party model trackers; it does not run the model or benchmark its media output. The parameter counts, context length, license, and API prices are consistent across multiple public sources, but exact release dates vary by a day between trackers, so "September 2026" is used here. Benchmark claims (including the Artificial Analysis score and the US$2.62M training cost) are vendor-reported and presented as indicative, not independently verified. Real media quality — character consistency, audio sync, artistic control — depends on prompting, fine-tuning, and downstream tooling not covered by the base model alone. ## What readers can do now - If you build content pipelines, prototype MiMo-V2.6 as the orchestration layer between specialized image and video models rather than a drop-in replacement. - Weigh self-hosting the open weights against API use: private assets favor hosting; sporadic use favors the per-token API. - For any client deliverable, keep a human in the loop on consistency and rights — open models raise capability, not automatic clearance. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-xiaomi-mimo-v26 Free to quote with attribution and a link to the original.