Overview
MiMo-V2.6-Flash is the efficiency tier of the MiMo-V2.6 series, released and open-sourced in the same batch as Pro. The architecture is likewise a sparse mixture of experts at 309 billion total and 15 billion activated parameters, with a 1M-token context, text, image, video and audio entering one model, and MIT-licensed weights.
The two models share one training recipe. Xiaomi describes this RL run as a large-scale expansion mixing coding, general agent work, vision and cybersecurity tasks. Flash completed 30 steps and roughly 750,000 trajectories in under six days at a reported RL cost of about 850,000 dollars. Xiaomi reports DeepSWE v1.1 rising from 48.8 to 65.7, OSWorld-Verified at 80.8 and AutomationBench at 52.3, all on the vendor side of the ledger.
What matters more directly for developers is price: 1 yuan input and 2 yuan output per million tokens, with cache hits at 0.02 yuan, again unchanged from the V2.5 series. It suits high-frequency, low-cost execution calls so that budget can be reserved for tasks needing longer reasoning chains.
Key Features
- Efficiency-tier configuration: A sparse MoE at 309B total and 15B active parameters, cutting per-token cost while keeping the 1M context.
- Native omni-modal input: Text, image, video and audio share the same input path as Pro.
- 1M-token context: Long documents and long tool traces can be handed over whole, without fragmenting context.
- Open-sourced with Pro: MIT-licensed weights, technical report and RL task environments all released for self-hosting.
- Low price tier: 1 yuan input and 2 yuan output per million tokens, 0.02 yuan for cache hits, aimed at concurrent execution calls.
- Shared RL recipe: Comes out of the same large-scale reinforcement-learning expansion as Pro, keeping the capability distribution aligned.
Use Cases
- High-frequency agent execution calls such as tool invocation, format conversion and batch file handling
- Long-document processing across contracts, reports and repositories into structured output
- Multi-modal batch work such as understanding and labeling image, audio and video content in bulk
- Budget-sensitive automation pipelines that route the bulk of calls through a low-cost tier
Pros
- A 15B-active sparse MoE brings unit cost down to the 1 yuan input tier, suited to high call volumes
- Shares one training recipe and modality path with Pro, so switching tiers is a small change
- Open weights under MIT, so it can be self-hosted
- Retains the 1M context in the efficiency tier instead of downgrading long tasks
Pricing
Token-based API pricing carried over from the V2.5 series: 1 yuan per million input tokens, 2 yuan per million output tokens, and 0.02 yuan for cache hits. Weights are open under MIT. Check official listings for current rates and quotas.
Summary
MiMo-V2.6-Flash is the efficiency tier of the MiMo-V2.6 series: a sparse MoE at 309B total and 15B active parameters that keeps the 1M-token context and native text, image, video and audio input, open-sourced alongside Pro from the same large-scale reinforcement-learning recipe. Xiaomi reports DeepSWE v1.1 rising from 48.8 to 65.7, OSWorld-Verified at 80.8 and an RL cost of about 850,000 dollars. Pricing carries over from the previous generation at 1 yuan input and 2 yuan output per million tokens.
Version History
- Xiaomi releases MiMo-V2.6-Flash as the efficiency tier for high-frequency execution (2026-09-22): MiMo-V2.6-Flash launched and was open-sourced alongside Pro as the native omni-modal efficiency tier of the series: a sparse MoE at 309B total and 15B active parameters with a 1M-token context and MIT-licensed weights. Xiaomi reports 30 RL steps, roughly 750,000 trajectories, under six days of training and about 850,000 dollars of compute. Vendor-reported results include DeepSWE v1.1 rising from 48.8 to 65.7, OSWorld-Verified at 80.8 and AutomationBench at 52.3. API pricing is unchanged from the previous generation at 1 yuan input and 2 yuan output per million tokens.