DeepSeek-V4.1-Flash Slashes Memory Costs for Long Contexts
DeepSeek has released DeepSeek-V4.1-Flash, a 552-billion-parameter model that slashes memory overhead to make running long-context AI agents significantly cheaper and more efficient.
DeepSeek launched DeepSeek-V4.1-Flash on September 10, a 552-billion-parameter Mixture-of-Experts model that drastically reduces the memory footprint required for long-context artificial intelligence operations. The model features a 40-layer Transformer backbone split into a 20-layer causal encoder and a 20-layer decoder, supporting text and image inputs up to a one-million-token context window. While auxiliary components can push its total parameter count to 763 billion, the model activates just 8 billion parameters on input and 16 billion on output, compared to the flat 13 billion parameters activated for both phases in the previous V4-Flash.
The defining achievement of the new architecture is its radical optimization of the key-value cache, which connects the prefill and decoding phases of inference. DeepSeek-V4.1-Flash reduces the global key-value cache to just 890 bytes per token, down from 3,514 bytes in V4-Flash and 48 kilobytes in the older V3.2. For a maximum context window of one million tokens, this slashes the memory requirement from 3.5 gigabytes to roughly 890 megabytes. Additionally, because local sliding-window states do not need to be saved to solid-state drives, persistent key-value cache storage is cut to about one-eighth of its predecessor's requirement.
For developers building long-running agents that accumulate massive histories of code, documentation, and tool outputs, these memory savings eliminate a major hardware bottleneck. The model achieves this efficiency without sacrificing capability. At maximum reasoning effort, DeepSeek-V4.1-Flash scores 40 on the Artificial Analysis Intelligence Index, placing it just behind Gemini 3.8 Flash High, which scores 41. However, the DeepSeek model achieves this at approximately one-quarter of the cost per task. Peak API pricing is set at $0.30 per million input tokens, $1.20 per million output tokens, and $0.006 per million cache-hit tokens. AI researcher Sebastian Raschka described the release as a "big overhaul" that represents a major leap forward in separating model capacity from serving costs.
This is our own summary of reporting by AlphaSignal
