Downloads · 30 days
10
25% of all-time downloads
yaya-3.6-FP8-dynamic
agentyaya/yaya-3.6-FP8-dynamic
yaya-3.6-FP8-dynamic is a image-text-to-text model from agentyaya. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
FP8-Dynamic (W8A8) quantization of agentyaya/yaya-3.6 — 35B-A3B hybrid Gated-DeltaNet + MoE vision-language model. ~36GB (from ~70GB BF16).
All-time downloads
40
Public
Parameters
35.1B
37.7 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors37.7 GB · 100%
Parameter types
How the weights are stored.
F8_E4M332.6B · 93%
Try a prompt
Base models
Model card
From the Hugging Face model README
yaya-3.6-FP8-dynamic
FP8-Dynamic (W8A8) quantization of agentyaya/yaya-3.6 — 35B-A3B hybrid Gated-DeltaNet + MoE vision-language model. ~36GB (from ~70GB BF16).
Produced with llm-compressor (FP8_DYNAMIC, data-free), following the official Qwen3-Next recipe. Kept in BF16 for quality: vision tower, MoE router gates, shared-expert gates, linear-attention (DeltaNet) modules, MTP, lm_head. Full recipe in recipe.yaml.
Serving
Loads directly in vLLM (compressed-tensors) — do not pass --quantization fp8:
vllm serve agentyaya/yaya-3.6-FP8-dynamic --trust-remote-code \
--max-model-len 262144 --tensor-parallel-size 2
Best throughput on Blackwell (B200) / Hopper FP8 tensor cores.
AI for indigenous and underrepresented communities.
- Task
- Image-Text-to-Text
- Type
- qwen3_5_moe
- License
- apache-2.0
- Created
- Aug 10, 2026
- Updated
- Aug 10, 2026