Modelpublic

esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T400

sourceHugging Faceapache-2.0updated 12d agoView on Hugging Face

0likes1.1kdownloads

Model Card

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T400

NVFP4 checkpoint of DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored, quantized for vLLM serving. NVFP4 here means W4A16 with FP8 scales, group size 16, weight only, 400 tensors. No calibration data was used.

HF repo: esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T400

At a glance

Field	Value
Format	`compressed-tensors` `nvfp4-pack-quantized`
Quantization	W4A16, group size 16, FP8 E4M3 scales, weight only, 400 NVFP4 tensors
Kept in BF16	lmhead, vision tower, MTP head, DeltaNet conv1d and inproj_a/b
Calibration	none
Base	`DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored` (Qwen3.8-27B, Apache 2.0)

Quantized set covers the MLP on all 64 layers, Q, K, V, O on the 16 full attention layers, the token embeddings, and the DeltaNet inprojqkv/z/outproj tensors. The lmhead stays at full precision on purpose: vLLM runs the output head unquantized regardless, and no vLLM release loads a compressed-tensors NVFP4 head, so quantizing it only breaks serving with zero runtime gain. Vision tower, MTP head, and the DeltaNet conv1d plus inproja/b dynamics stay at full precision following the JC1DA exception pattern.

Serving

Built for vLLM and engines consuming vLLM-compatible compressed-tensors checkpoints, following the JC1DA-style exception pattern. Verified serving: tensor parallel 2, 2048 context, 9 GB per GPU with ample KV headroom, coherent generation. Serve with --quantization compressed-tensors semantics (auto-detected from the config); FP8 KV remains a runtime flag, not checkpoint bytes. Keep vision intact and MTP for speculative decoding; no GGUF inside this repo.

Checkpoint

—Single model.safetensors about 19 GB
—config.json quantization_config.format=nvfp4-pack-quantized, quant_method=compressed-tensors, Qwen3_5ForCausalLM, 64 layers, hybrid GatedDeltaNet, 262144 context, MTP head
—tokenizer.json intact, chat template intact

Provenance

Derivative of DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored (Apache 2.0) which itself derives from Qwen/Qwen3.8-27B. The TWIN-TURBO tune targets reduced thinking tokens with matched output quality. Exception pattern follows the JC1DA INT4-W4A16 reference. A companion GGUF-parent checkpoint with the conservative 256-tensor recipe lives at esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4.

License

apache-2.0

Card written with AI assistance.