DeepSeek R1-0528 Qwen3 8B
DeepSeek post-trained Alibaba's Qwen3-8B-Base with R1-0528 reasoning traces. This is a trained adaptation, not a smaller set of experts extracted from the full R1 model.
Use the adaptation's tokenizer and configuration
The architecture matches Qwen3-8B, while the tokenizer configuration follows R1-0528. DeepSeek requires its own configuration files rather than the original Qwen package.DeepSeek-R1-0528-Qwen3-8B model card
A dense network has different deployment requirements
The configuration uses Qwen3ForCausalLM, 36 layers, 32 query heads and eight key-value heads. It specifies bfloat16 weights and fourfold YaRN scaling from 32,768 positions. It has no routed-expert structure like the 671-billion-parameter R1 network.R1-0528-Qwen3-8B checkpoint configuration
The upstream model is a pretrained foundation
Qwen identifies the base as a pretrained model with 8.2 billion parameters, including embeddings. DeepSeek supplies the reasoning adaptation. Its MIT release does not change the Apache 2.0 identification of the upstream base.Qwen3-8B-Base upstream model and licenseDeepSeek Qwen3-8B adaptation license
The small model accompanied the May R1 update
DeepSeek introduced this adaptation in its May R1 announcement. The publisher repository records its first model-file upload on 29 May 2025, one day after the parent model announcement. The available evidence supports May publication without a single confirmed first-release day.DeepSeek May 2025 release and deployment detailsDeepSeek Qwen3-8B publisher file history
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | May 2025 |
| Access | Downloadable |
| Model identifiers | DeepSeek-R1-0528-Qwen3-8Bdeepseek-ai/DeepSeek-R1-0528-Qwen3-8B |
| Base model | Alibaba Qwen3-8B-Base; dense 8B class.Qwen3-8B-Base upstream model and licenseDeepSeek-R1-0528-Qwen3-8B model card |
| Adaptation license | MIT.DeepSeek Qwen3-8B adaptation license |
| Upstream model license | Apache 2.0 for Qwen3-8B-Base.Qwen3-8B-Base upstream model and license |
| Configured position limit | 131,072; max_position_embeddings with YaRN scaling, not an independently verified service limit.R1-0528-Qwen3-8B checkpoint configuration |
Related models
- DeepSeek R1-0528
DeepSeek · DeepSeek R1
Updated full-size R1 reasoning model with more post-training computation, deeper reasoning and revised prompt handling.
Launch date Downloadable - DeepSeek R1
DeepSeek · DeepSeek R1
Reasoning language model that combines supervised examples and reinforcement learning on a large mixture-of-experts foundation.
Launch date Downloadable