Ornith-1.0-35B - a huginnfork Collection

huginnfork 's Collections

ThinkingCap-Qwen3.6-27B

Ornith-1.0-35B

updated 9 days ago

FP8 & NVFP4 quants of the 256-expert Ornith-1.0-35B MoE. Routers, SSM block and vision tower stay bf16. Serve in vLLM or SGLang.

huginnfork/Ornith-1.0-35B-FP8

Image-Text-to-Text • 35B • Updated Jun 30 • 161

Note FP8_DYNAMIC W8A8, 37.7 GiB (from 66 GiB bf16). KLD 0.0161 nats, PPL 6.7547 (+0.01 %) vs the bf16 base — essentially lossless.
huginnfork/Ornith-1.0-35B-NVFP4A16

Image-Text-to-Text • 21B • Updated Jun 30 • 21

Note NVFP4 W4A16, 23.3 GiB — a ~2.8x reduction. KLD 0.0600 nats, PPL 6.8990 (+2.15 %) vs base. Vision tower kept bf16, so no '!!!!' image bug.