DeepSeek R1-0528
R1-0528 updates the original R1 through additional post-training. The released weights retain the DeepSeek architecture; the small Qwen3 distillation is a separate model.
The prompt rules changed with the checkpoint
0528 accepts system prompts. It no longer requires a forced opening thinking tag.DeepSeek-R1-0528 model card
DeepSeek instructs existing deployments to replace both the checkpoint and tokenizer_config.json because tool-call formatting changed.DeepSeek May 2025 release and deployment details
Longer reasoning changes the output budget
In DeepSeek's AIME 2025 evaluation, average reasoning grew from about 12,000 to 23,000 tokens per question. Reported accuracy rose from 70% to 87.5%.DeepSeek-R1-0528 model card
The release added tools without interleaved thinking calls
The original 0528 API added function calls and JSON output, but could not call tools inside its thinking sequence. Its max_tokens setting counted thinking and final output together, up to 64K. These are historical release settings, not an offer of current R1 API access.DeepSeek May 2025 release and deployment details
Available weights outlasted the first-party API route
The old deepseek-reasoner alias later served other models and ended on 24 July 2026.DeepSeek dated release history
The 0528 repository still supplies downloadable weights.DeepSeek-R1-0528 model card
Its configuration records 163,840 maximum position embeddings, distinct from the advertised context.R1-0528 checkpoint configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | DeepSeek-R1-0528deepseek-ai/DeepSeek-R1-0528 |
| Context window | 131,072 tokens Published 128K context for downloadable weights; historical first-party service used 64K. DeepSeek May 2025 release and deployment details |
| Main model parameters | 671 billion total; 37 billion active per token, using the same V3-Base foundation.Original DeepSeek-R1 model cardDeepSeek May 2025 release and deployment details |
| Distributed parameter count | 685 billion including 14 billion in the multi-token prediction layer.DeepSeek May 2025 release and deployment details |
| Weight license | MIT.DeepSeek May 2025 release and deployment details |
Related models
- DeepSeek V4 Pro
DeepSeek · DeepSeek V4
Large text reasoning model with sparse attention, adjustable thinking effort and a separate general-availability release for agent tasks.
Launch date API · Downloadable - DeepSeek R1-0528 Qwen3 8B
DeepSeek · DeepSeek R1
Dense reasoning model trained from Qwen3-8B-Base using reasoning examples generated by the larger R1-0528 model.
Launch date May 2025Downloadable - 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