free open weights Engineers
DeepSeek V3
by DeepSeek
A Chinese open-weight model that delivers frontier-class performance at a fraction of the training cost.
5/5
Visit DeepSeek V3 - Context window
- 128K tokens
- Parameters
- 671B mixture-of-experts
- Open weights
- Yes
- Released
- Dec 2024
Pros
- + Frontier-class quality with open weights
- + Remarkably low training and inference cost
- + Strong coding and reasoning performance
Cons
- − 671B model needs substantial hardware to self-host
- − Hosted API raises data-residency questions for some
- − Smaller Western tooling ecosystem
What Is DeepSeek V3
DeepSeek V3 is an open-weight mixture-of-experts model from the Chinese lab DeepSeek. It made headlines by matching frontier-class quality while reportedly costing a fraction of what comparable models cost to train.
In Practice
DeepSeek V3 and its reasoning-focused siblings reset industry expectations about the cost of frontier AI. With open weights, it is widely self-hosted and served by inference providers, and it remains a benchmark for cost-efficient model design.