R1

DeepSeek R1

Quick answer: DeepSeek R1 is an open-weight chain-of-thought reasoning model released January 20, 2025 under the MIT licence. It is a 671B sparse MoE model (37B parameters active per forward pass) that achieves 96.4% on the AI2 Reasoning Challenge, rivalling closed frontier models at a fraction of the cost. Available via the DeepSeek API at ~$0.55 input / $2.19 output per 1M tokens, or free to self-host.

At a Glance

Where DeepSeek R1 leads

• 96.4% AI2 Reasoning Challenge — top-tier multi-discipline science reasoning, on par with the best closed models
• Open weights under MIT licence — free to download, audit, and self-host
• ~$0.55/$2.19 per 1M tokens via DeepSeek API — among the most cost-efficient frontier reasoning models
• 671B sparse MoE: high capability with only 37B active parameters per token
• Strong math and coding benchmark scores at launch (AIME 2024, MATH-500)

Where it lags

• Text only — no vision or multimodal capability
• 128K context window — smaller than Gemini 2.5 Pro (1M) or GPT-5.6 Sol (1.1M)
• Self-hosting a 671B MoE model requires significant infrastructure investment
• Performance on agentic/tool-use tasks not as thoroughly evaluated as closed models

Best for: Research, cost-sensitive reasoning pipelines, open-weight deployments where licence and auditability matter, and high-throughput scientific analysis tasks.

What DeepSeek R1 Is

DeepSeek R1 is DeepSeek's flagship reasoning model, trained using reinforcement learning from scratch to develop chain-of-thought reasoning capabilities without supervised fine-tuning as the primary training signal. This approach — described in the R1 technical report — produced a model that reasons through problems step by step before generating a final answer, similar in principle to OpenAI's o-series.

The model's architecture is a 671B sparse Mixture-of-Experts (MoE) transformer. MoE means only a fraction of the parameters (37B) are active for any given token, enabling high-quality outputs at lower inference cost than a dense 671B model would require. This efficiency makes R1 economically viable as an API offering and practical for teams with sufficient GPU infrastructure to self-host.

R1's release in January 2025 was significant not just for its performance — which matched or exceeded then-frontier closed models on math and science benchmarks — but for its licence: MIT, meaning unrestricted commercial use and the ability to fine-tune and deploy the weights without any API dependency.

Specifications

Pricing

Pricing per the DeepSeek API pricing page — rates may vary; verify before production use.

Context Window

DeepSeek R1 has a 128,000-token context window — roughly 90 pages of text in a single request. This is sufficient for most document analysis and coding tasks but is a constraint compared to long-context models like Gemini 2.5 Pro (1M) or GPT-5.6 Sol (1.1M). For long-context retrieval workloads, consider newer DeepSeek models or alternatives with larger windows.

Public Benchmark Scores

Scores from Benchgen evaluations or attributed third-party sources. The SHADE-Arena score of 0.0 indicates R1 did not engage in sabotage behaviour in that safety evaluation — this is expected for a well-aligned model.

DeepSeek R1 vs Alternatives

R1's strongest case vs o4-mini: marginally higher AI2 Reasoning Challenge score (96.4% vs 95.9%), MIT licence for unrestricted use, and lower API cost. The key trade-offs: R1 has a smaller context window (128K vs 200K) and is text-only (no vision), and o4-mini has more extensive agentic and tool-use evaluation coverage.

How DeepSeek R1 Performs on Real Agent Tasks

R1's SHADE-Arena score of 0.0 overall success reflects the benchmark's safety-evaluation design — not engaging in sabotage is the correct behaviour. Its AI2 Reasoning Challenge score of 96.4% — essentially matching o4-mini (95.9%) — confirms its strong multi-discipline reasoning capability.

For agent builders, R1's open weights are the primary differentiator. Teams that need to run inference on-premise (for data privacy, regulatory, or latency reasons), fine-tune the model for a specific domain, or avoid API dependency will find R1 uniquely positioned among frontier-class reasoning models. For cloud API use without those constraints, o4-mini and Gemini 2.5 Pro offer broader tool-use and multimodal evaluation coverage.

Use DeepSeek R1 via API

from openai import OpenAI  # DeepSeek API is OpenAI-compatible

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[{"role": "user", "content": "Solve this differential equation step by step..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: o4-mini · o3 · Claude 3.7 Sonnet · DeepSeek R1 0528
• Benchmarks: AI2 Reasoning Challenge · SHADE-Arena

Frequently Asked Questions

Specs and scores sourced from DeepSeek's official R1 technical report (January 2025) and Benchgen evaluations; third-party benchmark scores attributed inline. API pricing cited to the DeepSeek pricing page. Last updated 2026-07-23.