DeepSeek V4.1 Flash | Model library
DeepSeek V4.1 Flash is a 552B-parameter mixture-of-experts model that activates only 8B-16B parameters per token and supports a 1M-token context window for handling long documents and agentic workloads. The model cuts KV-cache memory usage by 4x compared to its predecessor while improving coding, reasoning, and agent performance.
- Activates just 8B parameters during input processing and 16B during generation, making it efficient despite the massive 552B backbone
- Reduces memory footprint by roughly 4x versus DeepSeek-V4-Flash-0731 through Compressed Sparse Attention 2 and a Causal Encoder-Decoder architecture
- Offers configurable reasoning-effort levels so developers can adjust latency and cost versus reasoning depth
- Handles multimodal input (text and images) natively with a 1M-token context window, designed for agents processing large amounts of data
DeepSeek-V4.1-Flash is a 552-billion-parameter mixture-of-experts model that activates only 8B parameters during input processing and 16B during generation. It handles images and text together, supports a 1-million-token context window, and uses a Causal Encoder-Decoder architecture with Compressed Sparse Attention 2 to keep long-context inference computationally reasonable. The model includes configurable reasoning-effort levels so you can dial up how much computation it spends thinking through problems, trading off latency and cost against answer quality.
Compared to the earlier V4-Flash version from July, this update cuts the KV-cache memory footprint by about 4× while improving performance on coding, reasoning, and agentic tasks. That combination of smaller memory demands and better results makes it practical for multimodal agents and workflows that need to process large inputs without melting your inference costs. The architecture seems built specifically for systems that need to handle long documents or conversations while staying efficient.
The code example shows straightforward integration through Baseten's API using an OpenAI-compatible client. You can enable reasoning with the `thinking` parameter and control reasoning depth with `reasoning_effort`, and the response separates out the model's internal reasoning tokens from its final answer. The usage tracking breaks down reasoning tokens separately, so you can see exactly what the model spent thinking versus generating.