DeepSeek R1 0528 Qwen3 8B

DeepSeek R1 0528 Qwen3 8B is a small chat model in DeepSeek's R1 reasoning family. Rather than being trained from scratch, it is a distillation: outputs and chain-of-thought traces from the larger R1 0528 release were used to post-train an 8-billion-parameter Qwen3 base model. That places it at the lightweight end of the R1 lineup — far smaller than the full R1 mixture-of-experts model, and intended for situations where the reasoning style of R1 is wanted at a size that fits on a single GPU. The model produces explicit reasoning before its final answer, which is the defining behavior of the R1 line and the main thing the distillation is meant to transfer. Its architecture and tokenizer come from the Qwen3 8B base, so it inherits that model's multilingual coverage and its serving characteristics under common inference stacks such as vLLM, SGLang, llama.cpp and Ollama. Because it is a dense 8B model, memory requirements are modest compared with the sparse R1 parent, and quantized builds are widely circulated. We do not currently track a verified context window, modality list, or benchmark throughput figures for this entry — our latency and tokens-per-second fields are unpopulated, so consult provider documentation for the serving limits that apply to a given endpoint. In practice this model is used where a reasoning-style answer is useful but the cost or hardware footprint of a large model is not justified: math and logic questions, code explanation and small refactors, structured extraction that benefits from a short deliberation step, and local or on-device experimentation. Compared with the full DeepSeek R1 0528, it trades depth on hard multi-step problems for much lower resource use; compared with a plain instruction-tuned 8B model, it spends more output tokens thinking before answering.