Qwen 2.5 3B — BonfyreFPQ v12 Native

Format: .fpq v12 (rANS entropy coded E8 + 6-bit tiles + FP16 scales)
Base model: Qwen/Qwen2.5-3B (3.09B params)
Size: 3.1 GB (vs 6.17 GB BF16 = 1.9× compression)
Bits per weight: 8.43 bpw
Quality: avg cosine 0.999714, worst 0.998316
Inference validated: logit cosine 0.9896, top-1 agreement 100% (Qwen2.5-3B)
Wan 2.1 video generation: successfully generated video from decoded v12 weights

Decode

./bonfyre-fpqx decode model-00001-of-00002.fpq  # outputs BF16 safetensors

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Base model

Finetuned

(369)

this model