AI Potluck
Model components / Fine-tuned / chat models

DeepSeek-R1

DeepSeek

Previous-generation reasoning-first model (January 2025) using large-scale GRPO RL to develop chain-of-thought before answering. 671B MoE plus 1.5B–70B distilled variants. Superseded by DeepSeek-R2 (April 2026), which inverts the architecture thesis; R2 is a 32B dense MIT model that runs on a single 24GB consumer GPU and reaches 92.7% AIME 2025. R1 retains 4.3M+ monthly downloads and remains the most-used open reasoning model.

DeepSeek-R1 (671B total / 37B active MoE, 128K context), reasoning model RL+SFT post-trained on DeepSeek-V3-Base (cold-start data + RL + SFT pipeline; R1-Zero is pure-RL variant); arXiv 2501.12948, released 2025-01-22. Verified live on HF June 2026.

Openness

3 high confidence
3.0
weights
open(MIT, on HF
data
closed(RL/SFT training data not released)
code
partial(inference
license
MIT(OSI-style permissive, no use restrictions)

MIT-licensed open weights with distillation rights, but RL/SFT post-training data and full training code are not released -> open_weights.

Adoption

4 high confidence
4.0

~5.58M downloads last month for the base R1 repo on HF, plus a very large family of R1 distills and the DeepSeek app/API surface. The R1 release was a landmark open-reasoning event; measured monthly downloads place it firmly in 1-10M, with broader app/API/derivative reach pushing toward the top of the band.

Capability

4 high confidence
4.0

Frontier open reasoning model at its Jan 2025 release (o1-class: matched/beat o1 on AIME, MATH-500, LiveCodeBench). As of June 2026 it is surpassed by DeepSeek-V3.2-Speciale/V4 and other 2026 frontier reasoners, so a 4 rather than a category-5 in mid-2026 terms. Figures are DeepSeek-reported on the HF card.

Unchanged since 2026-06-09 (last edited, not re-checked)