DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is the latest MIT-licensed release from DeepSeek, a natively multimodal mixture-of-experts model that accepts text and image input and returns text. Its causal encoder-decoder design pairs a 20-layer encoder with a 20-layer decoder, routing 6 of 384 experts per token for roughly 8 billion active parameters during prefill and 16 billion during decode, with a 196B-parameter Engram conditional memory accessed sparsely alongside the 552B backbone. Compressed Sparse Attention 2 and an FP4 KV cache of under 1KB per token make a full 1 million token context practical, while continuously controllable reasoning effort lets you trade depth for latency per request. Strong across coding, math, and agentic benchmarks, it is well-suited to long-horizon agent runs, repository-scale engineering, and document-heavy multimodal workloads.