MiMo-V2.6-Flash

The half-terabyte flagship’s little sibling, weights included. MiMo-V2.6-Flash is the efficiency checkpoint of the MiMo-V2.6 family: 309B total parameters with 15B active per token, sparse MoE, the full omni modality stack of the Pro model (text, image, video, audio - a 681M vision encoder plus audio tokenizer), 1M-token context, and a 5-layer multi-token-prediction speculative decoder. Weights are MIT on HuggingFace, shipping as a 177.7 GB FP8 checkpoint across 67 files, 39,625 downloads in the first week.

What runs where. The memory rule is total parameters resident, so 309B at 2-bit quantization is roughly 77 GB - the straddle class: too big for the 8 GB edge boxes, comfortable in a 96 GB workstation, at home in any 128 GB+ unified-memory or dual-GPU rig, where community GGUF builds (ggml-org’s build already at 10,840 downloads) land it. At FP8 as shipped, 177.7 GB means two datacenter-class GPUs or a node. Nothing here says “premier” about the hardware: this is the biggest open model that a rich single machine can hold at high compression.

Benchmarks, labeled. Vendor-run from the launch README: DeepSWE v1.1 67.9 (the Pro sibling 71.9, GPT-5.6 Sol 73.0, Claude Opus 5 74.0), Toolathlon-Verified 73.6, AutomationBench v1.0.6 52.3 - all within a few points of models four times its active size on the same rows, and all self-reported until third parties run them. The RL story matches the family: Flash trained under the same livestreamed reinforcement-learning program whose receipts Xiaomi published ($850K Flash / $2.62M Pro, under six days).

Where it sits. The family page’s own words: the 9B distill is the only member that approaches home hardware, and the Pro flagship is a two-node cluster model. Flash is the missing middle - frontier-adjacent quality, omni modalities, a published training bill, MIT weights, and a footprint that one serious desk can actually host. For buyers of the 96-128 GB class, this is the first MiMo worth putting on the shortlist.