MiMo-V2.6 Flash

// Overview

MiMo-V2.6 Flash is the efficiency-balanced checkpoint of Xiaomi's MiMo-V2.6 series, published to Hugging Face as MiMo-V2.6-Flash-RL on 21 September 2026 under an MIT licence. It keeps the series' full feature set — text, image, video and audio in one model, and a 1 million-token context — at 309 billion total parameters with 15 billion activated per token, against 1.02T/42B for its Pro sibling.

The smaller footprint comes from the backbone rather than from cut capabilities: 48 layers instead of 70, a hidden size of 4096 instead of 6144, 64 attention heads instead of 128, and 256 routed experts instead of 384 — with 8 activated in both cases. The vision encoder, audio stack and five-layer speculative decoder are shared with the Pro checkpoint, so the multimodal pipeline is identical.

Both checkpoints come out of the same training programme, which Xiaomi frames as scaling reinforcement learning toward self-improvement: a single mixed RL run over coding, general-agent, visual and cybersecurity tasks, fully asynchronous GRPO on batches of 1,568 prompts × 16 rollouts, and an agentic grader that ranks rollouts within each group rather than scoring them pass/fail.

On the comparison table Xiaomi published, Flash trails Pro by a few points on most agent and coding rows while actually leading it on CyberGym (95.1 against 94.0). The gap is widest on the harder security benchmarks — ExploitBench 25.3 against 47.9, SEC Bench Pro 47.5 against 66.3. Serving needs a fraction of the hardware: the documented SGLang command runs on a single node with tensor parallelism 8, and the vLLM command with tensor parallelism 4.

// Benchmarks

This model's scores

1. CyberGym95.1%
2. Terminal Bench 2.187.6%
3. OSWorld-Verified80.8%
4. MiMo Cyber Bench77.2%
5. Toolathlon-Verified73.6%
6. MiMo VisualCoding71.5%
7. DeepSWE v1.167.9%
8. MiMo Code Bench61.2%
9. JobBench61.2%
10. AutomationBench v1.0.652.3%
11. SEC Bench Pro47.5%
12. Terminal Bench 4.028.8%
13. Agents' Last Exam27.6%
14. ProgramBench26%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• MIT-licensed open weights at 309B total / 15B activated — the series' capabilities on hardware a single node can serve
• Same omnimodal input as the Pro checkpoint (text, image, video, audio) and the same 1M-token context
• Highest score in its own published comparison on CyberGym (95.1), ahead of MiMo-V2.6 Pro
• Single-node deployment documented for both SGLang (tensor parallelism 8) and vLLM (tensor parallelism 4), with EAGLE speculative decoding
• Shares the Pro checkpoint's 681M MiMo ViT and audio encoders, so multimodal behaviour carries over

// Best for

• Reach for it when you want the MiMo-V2.6 feature set on hardware that cannot hold a trillion-parameter checkpoint.
• Reach for it for long-context agent runs — repositories, tool traces, multi-session work — where the 1M window matters more than the last few benchmark points.
• Reach for it for multimodal pipelines that read screenshots, video or audio alongside text on self-hosted weights.
• Reach for it for vulnerability-discovery work in authorised environments, where it posts the series' best CyberGym score.

// How to access

// MiMo (open weights) — every version

The full lineage of the MiMo (open weights) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.