DeepSeek R1

DeepSeek R1 was introduced in early 2025 (arXiv 2501.12948) as DeepSeek's first dedicated extended-reasoning model: a 671B-parameter mixture-of-experts transformer with 37B parameters active per token that generates an internal chain of thought before its final answer, in the style of OpenAI's o1. Those reasoning tokens are billed as output. Its training pipeline was reinforcement-learning-heavy — Group Relative Policy Optimisation (GRPO) applied to reward verifiable outcomes in maths, coding and logic — rather than the supervised fine-tuning that dominates most instruction-tuned models.

R1 shipped under an MIT open-weights licence, so it can be self-hosted or reached through third-party hosts such as Together AI and OpenRouter at materially lower cost than DeepSeek's own API. It is a Chinese-lab model, which can carry data-residency or compliance constraints for some deployments. A later checkpoint, R1-0528, improved its vendor-reported SWE-bench Verified score from 49.2% to 57.6%.

The model's lifecycle is unusual and worth understanding before building against it. DeepSeek's own `deepseek-reasoner` API alias stopped pointing at R1 on 2025-08-21, when DeepSeek-V3.1 unified thinking and non-thinking modes behind one toggle; that alias now routes to deepseek-v4-flash's thinking mode and retires entirely on 2026-07-24. R1 has no standalone successor entry in this registry — reasoning capability now lives inside the hybrid V-series models. For a stable long-term target, the open-weights checkpoints (R1 and R1-0528) are the ones to pin.

Modelglass rates this entry's capability confidence as high on the reasoning dimension, but the broader ratings are qualitative rather than a single cited benchmark run, and the reasoning overhead can multiply effective cost three- to five-fold on hard problems — factor both in when comparing R1 against non-reasoning models on price.