---
license: other
base_model: SwissNeuron/Qwen3.8-27B-SwissNeuron-Derisked
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- qwen3.8
- qwen3_5
- fp8
- compressed-tensors
- multimodal
- vision
- reasoning
- image-text-to-text
- swissneuron
---

# Qwen3.8-27B-SwissNeuron-Derisked-FP8

FP8 dynamic quantization of [SwissNeuron/Qwen3.8-27B-SwissNeuron-Derisked](https://huggingface.co/SwissNeuron/Qwen3.8-27B-SwissNeuron-Derisked).

- Source revision: repaired direct-answer SFT + fresh latest-SFT DWM α=0.1
- Format: compressed-tensors W8A8 FP8 dynamic
- Weight quantization: static FP8
- Activation quantization: dynamic per-token FP8
- Gated-DeltaNet linear-attention modules, vision tower, MTP draft head, embeddings, and LM head remain BF16 for compatibility and stability
- Active config retains factor-4 YaRN to 1,048,576 tokens; validate extreme-context quality for your workload

## Runtime

Use a current vLLM/Transformers stack with Qwen3.8 (`Qwen3_5ForConditionalGeneration`) and compressed-tensors support. FP8 compute requires recent NVIDIA GPUs; Hopper, Ada, and Blackwell are suitable.

```bash
vllm serve SwissNeuron/Qwen3.8-27B-SwissNeuron-Derisked-FP8 \
  --tensor-parallel-size 1 \
  --trust-remote-code
```

## Thinking modes

The repaired checkpoint supports both modes through the included chat template:

```python
tokenizer.apply_chat_template(messages, enable_thinking=True, add_generation_prompt=True)
```

Set `enable_thinking=False` for lower-latency direct answers.

## Limitations

This is a quantized derivative. Re-evaluate critical workloads and long-context retrieval rather than assuming BF16 parity. The underlying model and release notes are documented in the source repository.