MiMo-V2.6-Flash

VentureBeat reports MiMo-V2.6-Flash launched alongside the Pro flagship as a smaller, cheaper sibling retaining the same 1-million-token context window and native text, image, audio, and video multimodality. Xiaomi prices Flash at roughly $0.14 per million input tokens and $0.28 per million output tokens via its API, making it among the cheapest frontier-tier options for high-volume production workloads. The piece frames Flash as a high-volume production counterpart to Pro, which scored 46 on Artificial Analysis' Intelligence Index to lead open weights. Both checkpoints are MIT-licensed, downloadable from Hugging Face, and runnable on rented or owned hardware, with Flash targeting cost-efficient deployments rather than peak intelligence.

MiMo-V2.6-Flash was open-sourced by Xiaomi under the MIT licence with Hugging Face weights posted September 21, 2026. The variant is a 309B-parameter sparse MoE with 15B active parameters, 48 layers mixing sliding-window and global attention, a hidden size of 4096, a 1-million-token context, and a 681M vision encoder plus two audio encoders for native multimodality. Flash includes a 5-layer multi-token-prediction drafter that proposes up to 7 tokens per speculative-decoding pass, and Xiaomi recommends SGLang for serving with vLLM tensor parallelism of 4. FP8 weights total 172.9 GB, and Xiaomi also released a 9B distilled checkpoint, a technical report, its RL code, and over 7,000 training environments alongside the models.

Xiaomi released and open-sourced the MiMo-V2.6 series on September 22, 2026, with MiMo-V2.6-Flash positioned as the efficiency-oriented variant balancing intelligence and cost alongside MiMo-V2.6-Pro. Flash is one of two natively omnimodal checkpoints that handle text, image, audio, and video in a single model. Xiaomi's own benchmark tables show MiMo-V2.6-Flash scoring 67.9 on DeepSWE v1.1, 26.0 on ProgramBench, 61.2 on the MiMo Code Bench, and 52.3 on Automation Bench, trailing MiMo-V2.6-Pro but exceeding MiMo-V2.5-Pro across coding and agentic evaluations. The release emphasizes scaling RL compute on verifiable, complex tasks to expand the capability frontier.