From the source
DeepSeek released the open weights for DeepSeek-V4.1-Flash, a 552B-parameter multimodal model built for coding agents, and
Baseten is serving it through its Model APIs.
- The model uses a Causal Encoder-Decoder architecture that activates 8B parameters per token for prefill and 16B for decode, trading a heavier decode for a lighter prefill than V4-Flash, which activates 13B for both.
- Baseten says this suits coding agents, where agentic loops generate far more prefill tokens than decode tokens.
- According to DeepSeek, the global KV cache needs just a quarter of the memory of V4-Flash's, helped by Compressed Sparse Attention 2 and FP4 KV caching.
- On coding and agentic benchmarks, Baseten says V4.1-Flash beats V4-Pro with roughly a third of the total parameters, though a 54.8 on Automation-Bench means it fails roughly half of complex workflows.
- It is DeepSeek's first non-experimental model with native image input, but a ZeroBench score below 50 still warrants human review of image-based reasoning.
- DeepSeek has retired the V4-Flash models on its platform and routes their traffic to V4.1-Flash, with V4-Pro traffic to be rerouted beginning September 14.





