DeepSeek's 1017 Model Uses 890 Bytes Per Token, but Fails 67 Percent of Fact Checks
The new revision slashes hardware requirements by 75 percent, yet independent tests show it invents answers in two-thirds of cases where it should have stayed silent.
By Maya Almasi ·
DeepSeek released revision 1017 of its V4.1 Flash model on October 23, 2026. The update reduces memory usage for the key-value cache to 890 bytes per token. This figure represents a 75 percent reduction compared to the previous V4 Flash version. For developers running long-term AI agents, this change lowers the hardware costs associated with storing context during extended tasks.
The model is a Mixture-of-Experts architecture with 552 billion total parameters. It activates 16 billion of those parameters during a single run. The system accepts both text and image inputs and supports a context window of 1 million tokens. DeepSeek describes the update as "wish-driven programming made real," a phrase that points to its intended use for open-ended coding tasks.
Independent testing by Artificial Analysis reveals a tradeoff in accuracy. On the Omniscience benchmark, the model answered only 43 percent of questions correctly. More critically, it hallucinated on 67 percent of questions where the correct response was to decline answering. The benchmark's main index scores the 1017 model at 5. By comparison, Kimi K3-max scored 20, while frontier proprietary models scored 42 or higher.
Despite the accuracy issues, the model performs well in efficiency metrics. DeepSeek data shows that the 1017 variant requires one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its cache compared to the standard V4 Flash. At off-peak rates, DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens. Cache-hit input costs drop to $0.003 per million tokens. Third-party providers often charge less due to the open-weight nature of the files.
The model ranks second among large open-weight systems on Artificial Analysis' intelligence index with a score of 45. The top-scoring model, Claude Opus 5.5, holds a score of 58. Xiaomi's MiMo V2.6-Pro is the only other model that beats the 1017 version on cost per task. The new release positions itself as a budget option for tasks where outputs can be verified, such as programming or format conversion, rather than for open-ended research where errors are harder to detect.
Source: Tom's Hardware
