XiaomiMiMo/MiMo-V2.6-Flash-RL model card

Fields

id

sources/hf-mimo-v2-6-flash-rl-card

URL

https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL

Publisher

Xiaomi MiMo

Kind

repo

Published

2026-09-21

Accessed

2026-10-18

notes

Fetched 2026-10-18. Weights repo created 2026-09-21 (HF API createdAt 2026-09-21T15:39:51Z), last modified 2026-10-09, MIT. Card model summary: sparse MoE, 309B total / 15B activated, 1M context, text+image+video+audio, 5-layer speculative decoder. Architecture table: 48 layers (39 SWA + 9 global), hidden 4096, SWA heads 64/8, global heads 64/4, head dims QK/V 192/128, sliding window 128, 256 routed experts with 8 activated, no shared experts. The card's own table is a vendor claim: Terminal Bench 2.1 at 87.6, and the HF eval-results widget repeats 87.6 for harborframework/terminal-bench-2.1. HF safetensors metadata: F32 12,032 + BF16 4,101,308,032 + F8_E4M3 3,859,808,256 + U8 302,795,194,368 = 310,756,322,688 parameters.

Notes

Fetched 2026-10-18. Weights repo created 2026-09-21 (HF API createdAt 2026-09-21T15:39:51Z), last modified 2026-10-09, MIT. Card model summary: sparse MoE, 309B total / 15B activated, 1M context, text+image+video+audio, 5-layer speculative decoder. Architecture table: 48 layers (39 SWA + 9 global), hidden 4096, SWA heads 64/8, global heads 64/4, head dims QK/V 192/128, sliding window 128, 256 routed experts with 8 activated, no shared experts. The card's own table is a vendor claim: Terminal Bench 2.1 at 87.6, and the HF eval-results widget repeats 87.6 for harborframework/terminal-bench-2.1. HF safetensors metadata: F32 12,032 + BF16 4,101,308,032 + F8_E4M3 3,859,808,256 + U8 302,795,194,368 = 310,756,322,688 parameters.

Cross-references

This record references no other records.