# DeepSeek R1 0528

DeepSeek R1 0528 is a reasoning model that develops chain-of-thought (CoT) capability through reinforcement learning rather than supervised reasoning traces. It shows strong math and code results on published benchmarks with fully open weights.

- **Model ID:** `deepseek/deepseek-r1`
- **Type:** chat
- **Providers:** deepinfra, bedrock
- **Context window:** 160,000
- **Maximum output tokens:** 16,384
- **Pricing:** $1.35/1M input tokens, $5.4/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/deepseek-r1

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'deepseek/deepseek-r1',
  prompt: 'Why is the sky blue?'
})
```

## About

DeepSeek R1 0528 was released January 20, 2025 and breaks from conventional reasoning model training. Instead of relying on human-written reasoning traces, DeepSeek applied reinforcement learning directly to the base DeepSeek-V3 weights. Unconstrained RL let emergent behaviors like self-verification, self-reflection, and long chain-of-thought generation develop organically.

The architecture is a 671B Mixture-of-Experts (MoE) model that activates 37B parameters per forward pass. On AIME 2024, DeepSeek R1 0528 achieves 79.8% Pass@1, on par with OpenAI o1. On MATH-500 it reaches 97.3%. The release documentation also highlights strong code and general reasoning performance.

The MIT License is permissive: many proprietary reasoning models impose stricter restrictions. DeepSeek released six smaller derivatives alongside the full model. The 32B and 70B versions match OpenAI o1-mini performance, giving teams cost-efficient alternatives to the full 671B model.

## What to consider

DeepSeek R1 0528 generates verbose reasoning traces before final answers. Budget output tokens generously and account for variable response length when estimating costs.

## When to use

### Best For

- Competitive mathematics, formal proof construction, and quantitative reasoning where AIME 2024 and MATH-500 benchmark results match your task
- Code generation, debugging, and algorithm design where RL-derived problem-solving patterns produce self-correcting chains before final output
- Complex multi-step reasoning in analytical domains (finance, science, engineering) where showing work and self-verification build trust in outputs

### Consider Alternatives When

- Your workload is primarily conversation, summarization, or content generation where extended reasoning traces add unnecessary output token cost
- You need hybrid thinking and non-thinking modes in one endpoint. DeepSeek-V3.1 or later supports both modes through the same model
- Latency is a hard constraint and you can't tolerate variable response times from long reasoning chains
- The task domain is purely creative writing or open-ended generation where structured reasoning adds no quality benefit

## Best for

- **Competitive mathematics:** Formal proof construction and quantitative reasoning where AIME 2024 and MATH-500 benchmark results match your task
- **Code generation and debugging:** Algorithm design where RL-derived problem-solving patterns produce self-correcting chains before final output
- **Complex analytical reasoning:** Multi-step reasoning in finance, science, and engineering where showing work and self-verification build trust

## Consider alternatives

- **Conversation or summarization:** Extended reasoning traces add unnecessary output token cost for content generation workloads
- **Hybrid thinking modes:** DeepSeek-V3.1 or later supports both thinking and non-thinking modes through the same endpoint
- **Strict latency requirements:** Variable response times from long reasoning chains are not acceptable when latency is a hard constraint
- **Pure creative writing:** Structured reasoning adds no quality benefit for open-ended generation tasks

## Frequently asked questions

### How was DeepSeek R1 0528 trained differently from other reasoning models?

DeepSeek applied reinforcement learning directly to the base model, bypassing the conventional step of training on human-written reasoning traces. Reasoning patterns like self-verification and reflection emerged from RL exploration rather than curated data.

### What are DeepSeek R1 0528's benchmark scores on mathematics?

79.8% Pass@1 on AIME 2024, on par with OpenAI o1 at release. On MATH-500 it scores 97.3%.

### What does the MIT License mean for using DeepSeek R1 0528 outputs commercially?

The MIT License permits commercial use. Many proprietary reasoning models impose stricter restrictions.

### What is the context window and architecture of DeepSeek R1 0528?

A context window of 160K tokens. The architecture is Mixture-of-Experts (MoE) with 671B total parameters, activating 37B per forward pass.

### When should I use DeepSeek R1 0528 versus DeepSeek-V3 or V3.1?

DeepSeek R1 0528 specializes in deep reasoning with extended chain-of-thought. DeepSeek-V3 and later variants are general-purpose models that balance reasoning with faster, lower-cost completions and suit mixed-workload deployments better.

### Does the reasoning trace appear in the API response?

Yes. The chain-of-thought trace appears in the response. This helps with debugging and with applications that display the model's reasoning to end users.

## Links

- [Model page](https://vercel.com/ai-gateway/models/deepseek-r1)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://deepinfra.com/deepseek-ai/DeepSeek-R1-0528)
- [Provider pricing](https://aws.amazon.com/bedrock/pricing/)