Activereasoningllm Open Source

DeepSeek R1

by DeepSeek· Released January 2025· Cutoff December 2024

DeepSeek R1 is a reasoning-focused large language model that excels in complex problem-solving, mathematics, and coding tasks. It uses a Mixture-of-Experts architecture with 671B total parameters (37B activated) and features a 128K token context window. The model is open-source under the MIT license and offers competitive pricing.

Official Site API Docs 🤗 Hugging Face 📄 Research Paper

Input cost

$0.55 per 1M tokens (cache hit), $2.19 per 1M tokens (cache miss)

Output cost

$8.00 per 1M tokens (cache hit), $8.00 per 1M tokens (cache miss)

Context window

128K tokens

Max output

—

Modalities

text

Parameters

671B (37B activated)

License

MIT

Capabilities

ReasoningCode GenerationMath Problem SolvingFunction CallingStreamingJSON Mode

Best For

Complex reasoning tasks, advanced mathematics, and code generation requiring deep logical analysis.

Strengths

Exceptional performance on math and coding benchmarks
Open-source with MIT license
Cost-effective pricing
Large context window of 128K tokens

Limitations

Not multimodal (text-only)
Slower inference compared to non-reasoning models
May overthink simple tasks
Limited to English and Chinese primarily

Use Cases

Solving advanced mathematical proofs

Generating complex code algorithms

Data analysis and logical reasoning

STEM tutoring and education

Research paper summarization

Competitive programming assistance

Scientific problem solving

Improvements Over Previous Model

Introduced chain-of-thought reasoning with reinforcement learning
Significantly improved math (AIME 2024: 79.8% vs 39.2%) and coding (Codeforces: 2029 Elo vs 1133 Elo) over DeepSeek V3
Open-source release under MIT license
Competitive pricing with cache-based discounts

Back to all models