DeepSeek V4.1 Flash: cheaper and more efficient, but prone to hallucination

DeepSeek updated its open-weight V4.1 Flash model (revision 1017): a 552B-parameter MoE with 16B active parameters and a 1M-token context. It cuts KV-cache HBM needs by 4x and SSD storage by 8x, but hallucinates on 67% of questions it cannot answer.

• 552B total parameters, 16B active, multimodal input and 1M-token context
• KV cache: 890 bytes per token, 4x less HBM and 8x less SSD than V4 Flash
• Pricing: $0.15 per 1M input and $0.6 per 1M output tokens, $0.003 on cache hits
• Hallucinates on 67% of unsupported questions; 43% correct answers