---
base_model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- qwen3_8
- multimodal
- compressed-tensors
- fp8
- speculative-decoding
- uncensored
language:
- en
---

# Qwen3.8 27B TWIN-TURBO Fable Cold Fusion 709-L Uncensored, FP8 + Calibrated FP8 KV

This is a calibrated FP8 derivative of [DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored](https://huggingface.co/DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored). The source checkpoint was pinned at revision `9cdc928b322203a58c362584882105a56918a640`.

The text model uses block-scaled FP8 weights and dynamic FP8 activations. Calibration also produced static tensor-wise FP8 KV-cache scales. The vision tower and restored MTP weights retain their source precision.

## Quantization

- **Toolchain:** LLM Compressor 0.13.0, compressed-tensors 0.18.0, Transformers 5.13.1, PyTorch 2.11.0+cu130
- **Text weights:** FP8, 128 x 128 block scaling
- **Activations:** dynamic FP8, group size 128
- **KV cache:** FP8 with static tensor-wise calibrated scales
- **Calibration data:** 512 UltraChat samples, maximum sequence length 2,048
- **Preserved from the source:** vision tower, MTP, embeddings, LM head, and linear-attention state projections
- **Format:** Transformers safetensors with compressed-tensors metadata

The exact serialized recipe is included in `recipe.yaml`.

## Chat templates

The release includes both the source model's custom template and the official Qwen template. Neither template was edited.

### DavidAU template: default text and tool path

`chat_template.jinja` is the exact template from the pinned DavidAU source revision. `chat_template-davidau-original.jinja` is an identical named copy. The source's other template variants are also included unchanged.

The DavidAU template preserves the model's custom TWIN-TURBO reasoning and tool controls. Text generation passed validation with this template.

### Official Qwen template: vision path

`chat_template-qwen-original.jinja` is the exact official template from [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B), pinned at revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`.

During validation, the DavidAU template rejected a user image with `System message cannot contain images.` The official Qwen template accepted the same image request and returned the correct answer. Use the official template when serving vision requests.

## vLLM examples

### DavidAU text and tool template

```bash
vllm serve <MODEL_ID> \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}'
```

### Official Qwen vision template

```bash
vllm serve <MODEL_ID> \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}' \
  --chat-template chat_template-qwen-original.jinja \
  --limit-mm-per-prompt.image 1
```

Set context length, concurrency, and memory allocation for the serving hardware. The commands above show the template and quantization-specific options only.

## Validation

Static checks:

- all 3 safetensor files and all 1,631 indexed tensors were readable and fully covered by the index;
- all 333 vision tensors exactly matched the source checkpoint;
- all 15 MTP tensors exactly matched the source checkpoint;
- 146 other protected tensors exactly matched the source checkpoint;
- 400 eligible text weights were FP8 and had 400 matching scale tensors;
- tokenizer and image/video processor contracts matched the source.

Runtime checks used vLLM 0.29.0 with FlashInfer 0.6.18:

- the multimodal target, calibrated FP8 KV cache, and MTP loaded successfully;
- the DavidAU template returned `QUANT_OK` over the text path with HTTP 200;
- the official Qwen template identified a solid red image as `red` with HTTP 200;
- MTP accepted tokens at both configured draft positions;
- no request error, queue wait, restart, or OOM occurred in the successful validation runs.

These are loading and canary checks, not a quality benchmark. Quantization can change model quality.

## Source model and license

The model behavior, training claims, and uncensoring method come from the DavidAU source model. Its pinned model card is preserved as `README.upstream.md`.

This derivative follows the source model's Apache 2.0 license. Review the source model card and license terms before use.