DeepSeek-V3

4.3

Based on 3 reviews

A highly efficient 671B mixture-of-experts model trained at a fraction of the cost of comparable models, proving that frontier AI does not require massive budgets. Delivers impressive coding and math performance rivaling GPT-4o. A game-changer for the open-source AI community with full weight availability.

Released

December 25, 2024

Parameters

671B (MoE, 37B active)

Context

128K

Pricing

Open Source

Benchmark Performance

Benchmark	Category	Score	Performance
MMLU	knowledge	88.5%	89
HumanEval	coding	82.6%	83
MATH	reasoning	90.2%	90
HellaSwag	language	88.9%	89
ARC-Challenge	reasoning	95.3%	95
Chatbot Arena ELO	overall	1334	76

Last updated: March 15, 2026

Benchmark scores may vary based on evaluation methodology and conditions.

Reviews(3)

Code With AI5h ago

Outstanding coding performance for an open-weight model. I run it locally and the quality rivals GPT-4o for most programming tasks. The MoE architecture keeps inference fast.

Data Cruncher5h ago

The efficiency of this model is staggering. Strong math and coding performance at a fraction of the compute cost. A game-changer for the open-source AI community.

AI Explorer5h ago

DeepSeek-V3 is remarkable for an open-source model. The fact that it competes with models costing 10x more to train is impressive. Great for self-hosting if you have the hardware.