DeepSeek v4.1 Flash (1017)

• 🧠 552B-parameter sparse Mixture-of-Experts; only 8B/16B active per token
• 📏 One-million-token context for long documents and agent traces
• 🆕 Causal Encoder-Decoder architecture: decoder KV projected from encoder states
• 👁️ Native vision encoder trained with text from pre-training onward
• ⚡ Served in FP8 quantization for efficient deployment
• 🔧 Tuned for coding, tool use, function calling and web search
• 🎯 Reasoning effort continuously adjustable from 1 to 100
• 🔒 Open weights on Hugging Face under the MIT license

DeepSeek

DeepSeek is a Chinese artificial intelligence company specializing in large language model development, founded in July 2023 by Liang Wenfeng. Based in Hangzhou, Zhejiang, the company is backed by High-Flyer, a prominent Chinese hedge fund also co-founded by…

DeepSeek V4.1 Flash is DeepSeek's efficiency-oriented multimodal Mixture-of-Experts model, and this catalog entry is the 1017 revision of that release. It is a sparse MoE model with a 552B-parameter backbone that accepts interleaved text and images and generates text, with a context window of roughly one million tokens.

The headline change versus DeepSeek V4 Flash 0731 and DeepSeek V4 Flash 0423 is architectural. V4.1 Flash is built on a Causal Encoder-Decoder design: 40 transformer layers split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states instead of computed layer by layer. That asymmetry means only about 8B parameters activate per input token and 16B per output token. Vision is also native here, with a from-scratch encoder trained alongside text rather than bolted on afterwards.

Compared with the much larger DeepSeek V4 Pro, V4.1 Flash trades total capacity for far lower per-token compute while keeping the million-token window. DeepSeek's own reported instruct-model evaluations include 90.9 on GPQA Diamond and 90.6 pass@1 on Terminal-Bench 2.1. Reasoning effort is continuously adjustable on a 1-to-100 scale, and the model supports function calling and web search. Weights are published openly under the MIT license. For a higher-capacity option, see DeepSeek V4 Pro 0813.