DeepSeek V4.1 Flash Reduces Memory Footprint by 75% in Latest Update

The landscape of large language models is currently divided between expensive proprietary systems and cheaper open-weight alternatives. DeepSeek has just released a revision of its V4.1 Flash model, designated as version 1017, which promises major efficiency improvements for developers running agentic workflows.

The updated model is a Mixture-of-Experts architecture with 552 billion total parameters, though only 16 billion are active during inference. It supports multimodal inputs including text and images, and it maintains a one million token context window. DeepSeek describes the release as enabling wish-driven programming, a nod to the emerging practice of defining tasks in broad, natural language terms rather than rigid code.

The most striking change is in resource consumption. Compared to the previous V4 Flash version, the 1017 update requires only one quarter of the high bandwidth memory for its key value cache. It also needs one eighth of the solid state drive storage, using just 890 bytes per token entry. This reduction significantly lowers the hardware barrier for self-hosting the model.

Pricing remains competitive for those using the API. DeepSeek charges 15 cents per million input tokens and 60 cents per million output tokens during off-peak hours. Cache hit inputs drop to 0.3 cents per million tokens. The company also offers a search feature for 75 cents per 1,000 queries, which is cheaper than its nearest rival, Parallel, at one dollar.

Performance benchmarks paint a mixed picture. Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index with a score of 45. It trails the top scoring Claude Opus 5.5, which holds a score of 58. However, the DeepSeek model is among the fastest in its category and only Xiaomi's MiMo V2.6-Pro is cheaper per task.

Reliability remains the primary concern. On the Omniscience benchmark, the model answered 43 percent of questions correctly. It hallucinated on 67 percent of questions where it should have declined to answer. Its main index score is 5, well behind Kimi K3 max at 20 and frontier models which score 42 or higher. This suggests the model is best suited for tasks with verifiable outputs like programming, while open ended research may still require human oversight or more expensive frontier systems.