DeepSeek's new model uses a quarter of the memory but fails most fact checks
The 1017 revision cuts storage requirements dramatically, yet it invents details on 67 percent of questions it should have declined to answer.
By David Almasi ·
DeepSeek has released a new version of its V4.1 Flash model, labeled 1017, which drastically reduces the hardware needed to run the software. The company describes the update with the slogan "wish-driven programming made real," suggesting the tool is built for open-ended tasks where users describe goals in broad terms rather than writing precise code.
The core technical change involves the model's memory usage. Compared to the previous V4 Flash version, the 1017 update requires only one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its context cache. Each token entry now takes up just 890 bytes. For a Mixture-of-Experts model with 552 billion total parameters, this reduction in resource consumption is a major shift for anyone looking to lower running costs.
DeepSeek positions the model as a cost-effective alternative to expensive commercial systems. On the company's own benchmarks, the 1017 model performs between GPT-5.6 Sol and Claude Opus 5. It even beats Anthropic's model in specific task tests. The company uses a cost-per-task chart based on the DeepSWE dataset to show that both standard and high-efficiency runs of the 1017 model land in the favorable top-left region of the performance graph.
Pricing reflects this efficiency focus. DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours. Cache-hit input drops to $0.003 per million tokens. Because the model weights are open, third-party providers can host the software, often undercutting the official API rates. The search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.00.
Independent testing by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It scores 45 on that metric, trailing the top-scoring Claude Opus 5.5, which holds a score of 58. The 1017 model is also among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro beats it on cost per task.
The main drawback is accuracy. On the Omniscience benchmark, the model answered 43% of questions correctly. This is a strong result for an open-weight system but falls behind proprietary options. More concerning, the model hallucinated on 67% of the questions it should have declined to answer. The benchmark penalizes wrong answers while ignoring non-answers. The 1017 model scored 5 on the main index, compared with 20 for Kimi K3-max and at least 42 for frontier models.
This reliability gap shapes where the model fits best. It works well for tasks with verifiable outputs, such as programming or format conversion, where an agent can check its work and retry. Open-ended research or precise document interpretation is riskier because errors are harder to catch. This pattern mirrors the launch of Kimi K3 several months ago, which was praised for being cheap but quickly fell out of favor after it invented facts during short lookups.
Open-weight models like the 1017 remain important for keeping AI agents affordable. They also serve as a check on price increases from frontier labs. Despite the hallucination rate, the 1017 model offers a specific trade-off: significantly lower infrastructure costs in exchange for a higher burden on the user to verify results.
