DeepSeek
We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities.
Running this yourself: likely needs a rented cloud gpu.
46.8
Quality Score
---
Arena ELO
33B
Parameters
16K
Context
This measures the amount of verifiable public evidence we have, not how capable the model is. A missing field means it has not been verified yet, not that its value is zero.
19 of 22 public signals
Use this section to answer one simple question first: how much outside evidence do we have that this model performs well? Structured benchmark scores appear first, then official provider evidence, then live arena signal.
This model has normalized benchmark rows, so scores here are directly comparable across benchmark sources.
Sign in to join the discussion
562.8K
Downloads
1.6K
Likes
Jan 2025
Released
5/5 signals
3/4 signals
4/5 signals
3/4 signals
4/4 signals
Parameters
32B
Training compute
Not reported
Dataset scale
Not reported
Base model
Qwen2.5-32B
Source-reported access: Open weights (unrestricted) · Speculative confidence
Gaps we are still tracking
These are recent benchmark or leaderboard claims from official provider sources. They are useful for freshness and context, but they are not treated the same as normalized independent benchmark rows.