DeepSeek V4.1 Flash Cuts Storage Needs by 80% in New Update
The updated model uses one-eighth of the SSD storage for its cache but still hallucinated on 67% of questions it should have declined.
By Jisoo Serrano ·
DeepSeek released version 1017 of its V4.1 Flash model, which requires only one-eighth of the SSD storage for its key-value cache compared to the previous version. The update also reduces High Bandwidth Memory usage to one-quarter of the original requirement, using just 890 bytes per token entry.
The V4.1 Flash is a Mixture-of-Experts model with 552 billion total parameters, but only 16 billion are active during a run. It accepts text and images and supports a context window of 1 million tokens. DeepSeek describes the release with the phrase "wish-driven programming made real," suggesting it is designed for open-ended tasks defined in broad terms.
In DeepSeek's internal benchmarks, the new model performs between GPT-5.6 Sol and Claude Opus 5, beating the Anthropic model in specific task tests. The company's cost-per-task chart shows that the 1017 variant completes tasks at rates similar to the maximum settings of more expensive competitors. At off-peak times, input costs $0.15 per million tokens and output costs $0.6, with cache-hit input dropping to $0.003.
Independent testing by Artificial Analysis ranks the 1017 model 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 has a score of 58. Only Xiaomi's MiMo V2.6-Pro beats the DeepSeek model on cost per task in that category.
Reliability remains a concern for the open-weight release. On the Omniscience benchmark, the model answered 43% of questions correctly. It hallucinated on 67% of the questions where it should have declined to answer. The benchmark scores the 1017 model at 5, while Kimi K3-max scores 20 and frontier models score 42 or more.
The model is best suited for tasks with verifiable outputs, such as programming or format conversion, where an agent can check its work. Open-ended research and fact lookups are better handled by frontier models or humans to avoid undetected errors. DeepSeek's search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.
The release continues a trend where capable open-weight models trade some reliability for lower running costs. While these models are important for making AI agents more affordable, they remain a check on price rises from major commercial labs. The 1017 update is now available for use through DeepSeek's API and can be run on third-party providers.
Source: Tom's Hardware
