Deepseek has released the V4.1 Flash (DSF) model, a multimodal Mixture-of-Experts (MoE) system featuring 552 billion total parameters, with 16 billion active during operation. The model supports a 1 million token context window and accepts both text and image inputs.
The V4.1 iteration improves efficiency over its predecessor, requiring one-quarter of the HBM and one-eighth of the SSD storage for its KV cache. It consumes 890 bytes per token entry, significantly reducing hardware requirements for deployment.
Pricing for the model is set at $0.15 per million input tokens and $0.6 per million output tokens during off-peak hours, with cache hits reducing input costs to $0.003. Search functionality is priced at $0.75 per 1,000 queries.
Despite its performance, the model faces challenges with reliability. Artificial Analysis reports that while the model achieves a 43% accuracy rate on questions, it exhibits a 67% hallucination rate on non-answerable queries, scoring 5 points on the Omniscience benchmark.
