DeepSeek V4.1 Flash

• 🧠 552B-parameter multimodal Mixture-of-Experts, only 8–16B parameters active per token
• 📏 One-million-token context window, extended from 64K training via YaRN
• 🆕 New Causal Encoder–Decoder design: 20 encoder plus 20 decoder layers
• 👁️ Native vision via DeepSeek-ViT encoder trained from pre-training onward
• ⚡ Prefill activates 8B parameters, decode 16B, aiding input-heavy agents
• 🔧 Tuned for code agents, function calling and web search workflows
• 🔒 MIT-licensed open weights, FP8 quantization on this deployment
• 🏢 Released September 2026 by DeepSeek, served as the "deepseek-flash" endpoint

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 the smallest member of DeepSeek's new architecture family and the company's cost-efficient multimodal tier, released in September 2026 under an MIT license. It is a Mixture-of-Experts model with 552B backbone parameters that activates roughly 8B parameters per input token and 16B per output token, and it handles up to one million tokens of context.

The clearest change from DeepSeek V4 Flash 0731 and the original V4 Flash 0423 is architectural. Where the V4 preview series used a conventional decoder stack with hybrid compressed/heavily-compressed attention, V4.1 Flash adopts a Causal Encoder–Decoder layout: a 20-layer causal encoder feeds a 20-layer decoder, and the decoder's global KV cache is projected from the final encoder states rather than each decoder layer's own hidden states. DeepSeek describes this as targeting a higher capability ceiling, faster inference and higher throughput while scaling to larger models.

Multimodality is also native here. Earlier Flash checkpoints were text models, with vision arriving as a separate experimental variant; V4.1 Flash instead trains a from-scratch vision encoder and MLP projector jointly with text from the start of language-model pre-training. Routing uses one shared plus 384 routed experts, six active per token.

Compared with DeepSeek V4 Pro and DeepSeek V3.2, V4.1 Flash trades total scale for sparsity, aiming at long-context agentic and coding workloads where input volume dominates cost. DeepSeek evaluates it on code-agent suites using its own harness at full 1M context.