tencent/Sequential-Hidden-Decoding-8B-n8-Instruct

1---2license: other3license_name: sequential-hidden-decoding4license_link: LICENSE5base_model:6- tencent/Sequential-Hidden-Decoding-8B-n87- Qwen/Qwen3-8B-Base8tags:9- sequential-hidden-decoding10- instruct11- text-generation12- conversational13---14 15# Sequential-Hidden-Decoding-8B-n8-Instruct16 17This is the instruction-tuned variant of **Sequential Hidden Decoding 8B n=8**, designed for conversational and instruction-following use cases.18 19- **Base model:** [Sequential-Hidden-Decoding-8B-n8](https://huggingface.co/tencent/Sequential-Hidden-Decoding-8B-n8)20- **Underlying architecture:** [Qwen3-8B-Base](https://huggingface.co/Qwen/Qwen3-8B-Base)21- **Scale:** 8x22- **Context Length:** 13107223- **Dtype:** bfloat1624 25## Key Idea26 27Sequential Hidden Decoding scales sequence length by preparing multiple Embedding matrices for the same token sequence, interleaving the results, and feeding the expanded sequence into the same Transformer. This model is the instruction-tuned release of the 8B n=8 variant.28 29## Serving (SGLang)30 31This model requires a patched version of [SGLang](https://github.com/sgl-project/sglang) for inference. See the [project page](https://github.com/Tencent/Sequential-Hidden-Decoding) for installation options.32 33```bash34python -m sglang.launch_server \35    --model-path tencent/Sequential-Hidden-Decoding-8B-n8-Instruct \36    --trust-remote-code \37    --tp-size 1 \38    --port 30000 --host 0.0.0.0 \39    --chunked-prefill-size -1 \40    --attention-backend fa3 \41    --mem-fraction-static 0.82 \42    --max-running-requests 32 \43    --context-length 131072 \44    --cuda-graph-max-bs 128 \45    --cuda-graph-bs 1 2 4 8 16 32 64 12846```47 48> **Note:** Sequential Hidden Decoding models process n×-length sequences internally, so `--chunked-prefill-size -1`, `--attention-backend fa3`, and conservative batch sizing are important for stability and performance.49 50## Chat Usage51 52This is an instruction-tuned model. Use the `/v1/chat/completions` endpoint:53 54```python55from openai import OpenAI56 57client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")58response = client.chat.completions.create(59    model="tencent/Sequential-Hidden-Decoding-8B-n8-Instruct",60    messages=[61        {"role": "system", "content": "You are a helpful assistant."},62        {"role": "user", "content": "Explain the idea of hidden decoding in simple terms."},63    ],64    max_tokens=512,65    temperature=0.7,66)67print(response.choices[0].message.content)68```69 70## Files71 72This repository includes the custom architecture files required by `trust_remote_code`:73 74- `configuration_qwen3_scale_seq.py`75- `modeling_qwen3_scale_seq.py`76 77## Related Models78 79| Model | Type | Notes |80|-------|:----:|-------|81| [Sequential-Hidden-Decoding-8B-n2](https://huggingface.co/tencent/Sequential-Hidden-Decoding-8B-n2) | Base | 2x scale base model |82| [Sequential-Hidden-Decoding-8B-n4](https://huggingface.co/tencent/Sequential-Hidden-Decoding-8B-n4) | Base | 4x scale base model |83| [Sequential-Hidden-Decoding-8B-n8](https://huggingface.co/tencent/Sequential-Hidden-Decoding-8B-n8) | Base | 8x scale base model |84| [Sequential-Hidden-Decoding-8B-n8-Instruct](https://huggingface.co/tencent/Sequential-Hidden-Decoding-8B-n8-Instruct) | Instruct | Instruction-tuned 8x scale model |85 86## Citation87 88```bibtex89@article{hidden_decoding_2026,90  title   = {Hidden Decoding: Scaling Sequence Length in Pretraining},91  year    = {2026},92  url     = {https://welm.weixin.qq.com/posts/hidden_decoding/}93}94```95 96## License97 98This model is released under the [License Terms of Sequential-Hidden-Decoding](LICENSE).99