Downloads · 30 days
71
100% of all-time downloads
sdworld/MiMo-V2.6-Flash-SD
MiMo-V2.6-Flash-SD is a text generation model from sdworld. Use it when you need the model to write or continue text. It is set up for sglang. The card lists the license as mit.
A compressed build of MiMo-V2.6-Flash-RL (309B MoE, 15B active) that runs on one DGX Spark (GB10, 128 GB unified memory) with sglang, all experts resident, 256k context. Text-only (the vision and audio towers are not…
Downloads · 30 days
71
100% of all-time downloads
All-time downloads
71
Public
Parameters
83.7B
98.2 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors95.6 GB · 97%
How the weights are stored.
U875.7B · 90%
From the Hugging Face model README
A compressed build of MiMo-V2.6-Flash-RL (309B MoE, 15B active) that runs on one DGX Spark (GB10, 128 GB unified memory) with sglang, all experts resident, 256k context. Text-only (the vision and audio towers are not included).
Measured on one DGX Spark (single stream, greedy, 400-token generations): prose 28.8 / code 38.2 / math 38.8 tok/s; 51.8 tok/s aggregate with 3 concurrent streams; GSM8K greedy 188/200.
model_pp0_*.safetensors, model_mtp.safetensors, model.safetensors.index.json, config.json, tokenizer filesdense_codes/ and sidecars/: auxiliary files the launcher loadsThe launcher, the patched sglang files and the benchmark scripts are in the companion GitHub repository
(MiMo-V2.6-Flash-SD). Download this repository with
pip install -U huggingface_hub
hf download sdworld/MiMo-V2.6-Flash-SD --local-dir ./MiMo-V2.6-Flash-SD
and point the launcher at it (CKPT=./MiMo-V2.6-Flash-SD ./run.sh).
MIT, as the base model.