DeepSeek-R1-0528-Qwen3-8B

// Overview

DeepSeek-R1-0528-Qwen3-8B is an 8-billion-parameter open-weight reasoning model from DeepSeek, released on 29 May 2025 alongside the larger DeepSeek-R1-0528 update. DeepSeek built it by distilling the chain-of-thought reasoning from R1-0528 into Qwen3-8B Base — in other words, it takes the reasoning traces produced by the big 685B R1-0528 model and uses them to post-train Alibaba's small Qwen3-8B base model. The result is a compact model that thinks step by step before answering.

Its network architecture is identical to Qwen3-8B, so it runs with the same inference setup, but it ships with DeepSeek's own tokenizer configuration rather than Qwen's. DeepSeek recommends a sampling temperature of 0.6, supports system prompts, and (unlike some earlier R1 distills) does not require you to manually prepend a thinking token to trigger reasoning. The whole model is released under the permissive MIT License, which allows commercial use and further distillation.

DeepSeek positions DeepSeek-R1-0528-Qwen3-8B as a way to bring R1-grade reasoning to commodity hardware. On its own model card, DeepSeek reports that it reaches state-of-the-art results among open-source models on AIME 2024 — beating the original Qwen3-8B by 10 points and roughly matching the much larger Qwen3-235B-thinking on that math test.

// Benchmarks

1. AIME 202486%
2. AIME 202576.3%
3. HMMT Feb 202561.5%
4. GPQA Diamond61.1%
5. LiveCodeBench (2408-2505)60.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• State-of-the-art math reasoning for an 8B open model — 86.0 on AIME 2024 per DeepSeek's model card, +10 points over Qwen3-8B.
• Small enough to run locally on a single consumer GPU (or quantized on a laptop), bringing R1-style chain-of-thought to commodity hardware.
• Fully open under the MIT License, with weights on Hugging Face, so it can be self-hosted, fine-tuned, and distilled commercially.
• Drop-in Qwen3-8B inference compatibility (same architecture), with system-prompt support and no manual think-token needed.

// Best for

• Local and on-device reasoning where a hosted frontier model is impractical or privacy-sensitive.
• Math and competition-style problem solving (AIME / HMMT) on a small footprint.
• Cost-controlled coding and logic tasks via self-hosting under an open license.
• A base for further fine-tuning or distillation experiments under MIT.

// How to access

// DeepSeek R1 Distill — every version

The full lineage of the DeepSeek R1 Distill line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.