The open-weights model family from DeepSeek, known for strong reasoning on math and code. Running in both Berges layers.
DeepSeek is a Chinese AI lab whose open-weights models have a reputation for punching above their weight. The DeepSeek family uses mixture-of-experts architecture with notably efficient training, and the resulting models are particularly strong on reasoning, math, and code.
Berges AI runs DeepSeek in both layers. The lightweight variant (DeepSeek-V4-Flash) handles the 90% of questions that don't need deep thought. The larger Pro variant escalates for harder ones.
DeepSeek publishes under an open license, and the training methodology is documented in detail. That's rare in this space and part of why the model is interesting.
DeepSeek is one of the strongest open-weights models on technical tasks. Code review, debugging, math problems, structured reasoning. All things it does well.
DeepSeek's training process is famously efficient. That translates downstream to faster, cheaper inference, which is part of why we use it for the fast layer.
DeepSeek publishes detailed technical reports about how the models are built. The weights are open and the architecture is public.
DeepSeek publishes weights openly, and the same models run on many open-weights hosts. Berges AI provides one chat experience around them, with encryption at rest and no training on your conversations.
DeepSeek ships a small/fast variant and a larger reasoning variant. We use both: Flash for everyday questions, Pro for harder ones, with the cascade deciding when to escalate.
On math, code, and many reasoning tasks, the gap with frontier closed models is small. On general everyday tasks, also small. On frontier capability, OpenAI is still ahead. DeepSeek is excellent for its weight class.