Downloads · 30 days
547
27% of all-time downloads
esatapedico/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T336
Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NVFP4-W4A16-T336 is a text generation model from esatapedico. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
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 w…
Downloads · 30 days
547
27% of all-time downloads
All-time downloads
2K
Public
Parameters
27.8B
64.7 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors23 GB · 100%
How the weights are stored.
U822.6B · 82%
From the Hugging Face model README
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
| 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.
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.
model.safetensors about 21 GBconfig.json quantization_config.format=nvfp4-pack-quantized, quant_method=compressed-tensors, Qwen3_5ForCausalLM, 64 layers, hybrid GatedDeltaNet, 262144 context, MTP headtokenizer.json intact, chat template intactDerivative 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.
I post updates on new conversions, benchmarks, and what I'm working on over on my site. Follow along there to keep up with new releases and the work in progress. If you'd like to support more of it, you will find the options there. I do this on consumer hardware and like seeing how far it goes. More is on the way.
🌐 esatapedico.com. Updates, work-in-progress, and ways to support.
apache-2.0
Card written with AI assistance.