DeepSeek‑R1‑0528: How the New Open‑Source LLM Outperforms Gemini and Claude

DeepSeek‑R1‑0528 Open‑Source Release

DeepSeek‑R1‑0528 model weights have been uploaded to HuggingFace (project link). The model, trained on DeepSeek‑V3‑0324 with 660B parameters, shows a dramatic evolution in coding ability and longer reasoning time.

Benchmark results: On LiveCodeBench, DeepSeek‑R1‑0528 performs on par with o3‑mini (High) and o4‑mini (Medium), surpassing Gemini 2.5 Flash.

Key highlights (summarized by the community):

Deep reasoning comparable to Google models

Improved text generation – more natural and better formatted

Unique inference style – fast yet thorough

Extended thinking time – single‑task processing up to 30‑60 minutes

Long‑thinking capability is a major discussion point; some users observed reasoning times exceeding 25 minutes.

Programming performance: Users report that DeepSeek‑R1‑0528 can generate correct code on the first try, outperforming Claude 4 and Gemini 2.5 Pro in front‑end and back‑end tasks.

Comparisons with Gemini 2.5 Pro on research prompts, live source searches, and AI SaaS strategy show mixed results, often a tie.

Additional tests demonstrate stable, fast reasoning for tasks such as 3D animation generation, website design, physics simulations, and classic puzzles like the farmer‑fox‑goose‑beans problem.

The model also handled complex relational queries and lengthy reasoning chains without interruption, indicating improved compute resources after release.

Overall, DeepSeek‑R1‑0528 represents a significant step forward for open‑source LLMs, delivering strong coding, reasoning, and multi‑step problem‑solving capabilities.