DeepSeek 1017 Cuts Memory Costs by 75 Percent but Fails 67 Percent of Fact Checks
The new open-weight model reduces hardware requirements dramatically, yet independent testing reveals it frequently invents information on questions it should have skipped.
By Chloe Zhao ·
DeepSeek released a revised version of its V4.1 Flash model on Thursday, labeled 1017. The update significantly reduces the memory required to run the software, but independent tests show the model still struggles to distinguish between facts and invented details.
The 1017 model requires only 890 bytes of memory per token entry. This figure represents a 75 percent reduction compared to the previous V4 Flash version. The model also uses one-eighth of the solid-state drive storage for its key-value cache. DeepSeek, a Chinese developer, describes the model as using 552 billion total parameters, with 16 billion active during any single run.
The company positions the 1017 update as a tool for broad, open-ended tasks. Its product page features the slogan "wish-driven programming made real." DeepSeek lists off-peak input costs at $0.15 per million tokens and output costs at $0.60. Cache-hit input drops to $0.003 per million tokens. The search function costs $0.75 per 1,000 queries, a rate lower than its closest rival, Parallel, which charges $1.
Independent analysis from Artificial Analysis places the 1017 model in a mixed position. It ranks second among large open-weight models on an intelligence index with a score of 45. That trails the top-scoring Claude Opus 5.5 model, which holds a score of 58. However, the 1017 model is among the fastest in its category and offers a lower cost per task than most competitors.
The primary weakness is reliability. On the Omniscience benchmark, the model answered only 43 percent of questions correctly. It hallucinated on 67 percent of questions it should have declined to answer. The benchmark's main index scores the 1017 model at 5. That compares with a score of 20 for Kimi K3-max and scores of at least 42 for leading proprietary models.
Experts suggest the model works best for tasks with verifiable outputs, such as programming or format conversion. In those cases, an agent can check the work and retry if errors appear. Open-ended research or precise document drafting remains risky because hallucinations are harder to detect.
The release continues a trend where capable open-weight models trade accuracy for lower running costs. This approach allows developers to run AI agents more cheaply while providing a potential check on pricing from large commercial labs.
Source: Tom's Hardware
