New Deepseek update slashes memory costs but still makes things up

The competitive landscape for large language models remains sharply divided between expensive proprietary systems and cheaper open-weight alternatives. Deepseek has just released a new revision of its V4.1 Flash model, designated 1017, which promises significant efficiency gains while retaining the known weaknesses of its open-source peers.

This Mixture-of-Experts model features 552 billion total parameters, with only 16 billion active during processing. It handles both text and images and supports a one million token context window. The company markets the release with the phrase wish-driven programming made real, suggesting it is designed for broad, open-ended tasks. However, the technical focus is on reducing the hardware required to run the model effectively.

Efficiency is the headline feature of this update. Compared to the previous V4 Flash version, the new model requires only one-quarter of the high bandwidth memory and one-eighth of the solid state drive storage for its cache. Each token entry now uses just 890 bytes. Deepseek provided a cost-per-task chart based on the DeepSWE September 2 dataset to support these claims. The results show that even low-effort runs of the 1017 version complete tasks at rates similar to the maximum settings of more expensive competitors.

Pricing remains aggressive for users who prefer open weights. At off-peak times, input costs are 15 cents per million tokens, while output costs are 60 cents. Cache-hit input drops to 0.3 cents per million tokens. Because the weights are open, third-party providers can host the model and often offer lower prices than Deepseek's own API. The search feature is also cheaper than rivals, costing 75 cents per 1,000 queries compared to 1 dollar from the closest competitor.

Performance benchmarks place the model in a strong position. Artificial Analysis ranks it second among large open-weight models on its intelligence index with a score of 45. This trails the top-scoring Claude Opus 5.5, which earned 58 points. The model is among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro beating it on cost per task. Deepseek's internal tests also show it outperforming Anthropic's Claude Opus 5 in specific task tests.

Reliability remains the primary drawback. On the Omniscience benchmark, the model answered 43% of questions correctly, a solid result for open-weight models but behind proprietary systems. The bigger issue is that it hallucinated on 67% of questions it should have declined to answer. Its overall index score is 5, significantly lower than the 20 for Kimi K3-max and the 42 or more for frontier models.

This pattern suggests the model is best suited for tasks with verifiable outputs, such as programming or format conversion where errors can be caught and corrected. Open-ended research or precise document drafting still carry a high risk of unnoticed errors. This fast-and-loose behavior mirrors earlier launches like Kimi K3, which was praised for low costs but criticized for inventing facts. Despite these flaws, open-weight models continue to play a vital role in keeping AI agents affordable and providing a check on frontier lab pricing.