DeepSeek 1017 Reduces Memory Use by 75 Percent but Hallucinates in 67 Percent of Cases
The new revision of the open-weight model cuts hardware storage requirements dramatically, yet independent tests show it frequently invents facts when it should decline to answer.
By Tyler Khalil ·
DeepSeek released a new revision of its open-weight V4.1 Flash model, designated 1017, on October 23, 2026. The update promises a sharp reduction in memory requirements, but independent testing reveals a high rate of factual errors that may limit its practical use for some applications.
The efficiency gains are substantial for operators managing large-scale inference. V4.1 Flash 1017 requires only 890 bytes per token entry in its KV cache. This represents a 75 percent reduction compared to the previous V4 Flash version. The new model also needs one-quarter of the High Bandwidth Memory and one-eighth of the solid-state drive storage for that cache. These reductions lower the hardware costs associated with running long-running agentic workloads.
DeepSeek positions the model as suitable for broad, open-ended tasks, using the phrase wish-driven programming made real on its product page. This language suggests an alignment with vibe coding, where users describe goals in general terms rather than providing precise instructions. The model is a Mixture-of-Experts architecture with 552 billion total parameters, of which 16 billion are active during a run. It accepts text and images and supports a context window of 1 million tokens.
Cost efficiency remains a central advantage. DeepSeek prices input tokens at $0.15 per million and output tokens at $0.60 during off-peak hours. Cache-hit input costs drop to $0.003 per million tokens. The company also charges $0.75 for 1,000 search queries, a rate lower than its closest rival, Parallel, which charges $1. Because the weights are open, third-party providers can host the model and often undercut DeepSeek's own API pricing.
Performance benchmarks show the model competing with top commercial systems. 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 scores 58. In DeepSeek's own tests, the model performs between GPT-5.6 Sol and Claude Opus 5, beating Anthropic's model in specific task tests. The model is among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro offering a lower cost per task.
Reliability remains the primary drawback. On the Omniscience benchmark, the model answered only 43 percent of questions correctly. More notably, it hallucinated on 67 percent of questions it should have declined to answer. The benchmark penalizes wrong answers while not penalizing non-answers, resulting in a score of 5 for the 1017 model. This compares to a score of 20 for Kimi K3-max and scores of 42 or higher for frontier models.
These results suggest the model is best suited for tasks with verifiable outputs. Programming, optical character recognition with review, and format conversion allow agents to check their work and retry. Open-ended research, fact lookups, and drafting precise documents are better handled by frontier models or humans, as hallucinations in these areas are difficult to catch. This pattern of capable but loose behavior mirrors earlier issues with Kimi K3, reinforcing the trade-off between cost and reliability in the open-weight sector.
