license: apache-2.0 base_model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored pipeline_tag: text-generation library_name: transformers tags:
- nvfp4
- qwen3.8
- qwen3.5
- blackwell
- mtp
- speculative-decoding
- turbo
- twin-turbo
- fable
- cold-fusion
- compressed-tensors
- vllm language:
- en
- multilingual quantization_config: quant_method: compressed-tensors format: nvfp4-pack-quantized
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 | lm_head, 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, Q, K, V, O on the 16 full attention layers, the token embeddings, and the DeltaNet in_proj_qkv/z/out_proj tensors. The lm_head 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 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, 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.safetensorsabout 19 GB config.jsonquantization_config.format=nvfp4-pack-quantized,quant_method=compressed-tensors,Qwen3_5ForCausalLM, 64 layers, hybrid GatedDeltaNet, 262144 context, MTP headtokenizer.jsonintact, 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.