Sign up

esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T336

🤗 Hugging Face sourcetext-generationapache-2.027.3B params32 GBsafetensors✓ HF checksums availableupdated today
No torrent yet →

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T336

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, 336 tensors. No calibration data was used.

HF repo: esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T336

At a glance

Field Value
Format compressed-tensors nvfp4-pack-quantized
Quantization W4A16, group size 16, FP8 E4M3 scales, weight only, 336 NVFP4 tensors
Kept in BF16 lm_head, full-attention q/k/v/o, token embeddings, vision tower, MTP head, DeltaNet conv1d and in_proj_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 (192 tensors) plus the DeltaNet in_proj_qkv/z/out_proj tensors on the 48 linear-attention layers (144 tensors). Full-attention projections and token embeddings stay at full precision, where precision buys the most quality; the lm_head stays at full precision because vLLM runs the output head unquantized regardless. Vision tower, MTP head, and DeltaNet conv1d plus in_proj_a/b dynamics stay at full precision following the JC1DA exception pattern.

Serving

Built for vLLM and engines consuming vLLM-compatible compressed-tensors checkpoints. Verified serving: tensor parallel 2, 2048 context, coherent generation. Measured perplexity on a diverse English payload is on par with the 256-tensor parent checkpoint (3.19 vs 3.20), so the DeltaNet quantization costs nothing measurable. 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 21 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. Sibling checkpoints: the conservative 256-tensor GGUF parent at esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4 and the 400-tensor serving build at esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T400.

License

apache-2.0

Card written with AI assistance.