DeepSeek-V4.1-Flash

What it is

DeepSeek-V4.1-Flash is a multimodal large language model from DeepSeek, released on 10 September 2026 under the API model name deepseek-flash ^[1]. DeepSeek described it as the smallest model in its new architecture family, with native visual understanding ^[1]. It replaced V4-Flash and V4-Flash-Vision-Exp ^[1].

Architecture

The model has 552 billion parameters and uses a mixture-of-experts design. Its most distinctive feature is what DeepSeek calls a causal encoder-decoder architecture. In this design, 8 billion parameters are active for input processing (prefill) and 16 billion for output generation (decode) ^[1][2]. Prefill is the step in which the model reads the prompt, and decode is the step in which it writes the answer one token at a time.

The network has 40 layers, made up of a 20-layer encoder and a 20-layer decoder. Each MoE layer has one shared expert and 384 routed experts, and 6 routed experts are active for each token ^[2].

The model card lists further components: Compressed Sparse Attention 2, FP4 KV caching, single-pass mHC, a 196-billion-parameter Engram conditional memory, DSpark speculative decoding and a DeepSeek-ViT vision encoder ^[2]. Speculative decoding is a technique in which a fast draft predicts several tokens that the main model then checks.

The KV cache needs a quarter of the HBM (high-bandwidth memory) and an eighth of the SSD storage of the previous generation, at about 890 bytes per token ^[1][2].

Training

The model was pre-trained on 45 trillion tokens of multimodal data, with its context extended to 1M tokens. Post-training used supervised fine-tuning, then reinforcement learning, then on-policy distillation. Reasoning effort is adjustable from 1 to 100 ^[2].

Results

For the Instruct version at maximum effort, the model card reports 90.9 on GPQA Diamond, 36.8 on HLE, a Codeforces rating of 3471, 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1 ^[2]. DeepSeek said its benchmark results were ahead of flagship models, including DeepSeek-V4-Pro ^[1].

Release and later updates

Weights and a technical report were published on Hugging Face under the MIT License ^[1][2]. From 14 September 2026 the model also served requests sent to deepseek-v4-pro, until a V4.1-Pro launches ^[1]. An update, DeepSeek-V4.1-Flash-1017, followed on 17 October 2026 ^[3].