DeepSeek's 1017 Model Uses 890 Bytes of Memory Per Token
The revision cuts hardware needs by 75 percent, but tests show it invents facts in 67 percent of cases where it should have declined to answer.
By Natalia Costa ·
890 bytes. That is the amount of memory DeepSeek's latest V4.1 Flash 1017 model requires for each token entry in its KV cache. This figure represents a sharp drop from previous versions, which demanded significantly more storage to perform the same computational tasks. The company released the update on October 23, 2026, positioning it as a key step in making large language models cheaper to run for developers and businesses.
The V4.1 Flash is a Mixture-of-Experts architecture with 552 billion total parameters. Only 16 billion of those parameters are active during any given run. The model supports text and images and handles a context window of 1 million tokens. Compared to the earlier V4 Flash, the 1017 revision needs one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive space. This reduction in hardware requirements allows the model to operate on less expensive infrastructure.
DeepSeek claims the model performs competitively with top-tier commercial systems. In its own benchmarks, the 1017 sits between GPT-5.6 Sol and Claude Opus 5. It actually beats Anthropic's model on certain task tests. The company highlights cost efficiency as the primary benefit. Using data from the DeepSWE September 2 dataset, both standard and 1017 variants complete tasks at rates comparable to the high-cost settings of rival models. DeepSeek's API pricing reflects this efficiency, with input costs at $0.15 per million tokens and output at $0.6 per million. Cache-hit input drops to just $0.003.
Independent testing by Artificial Analysis reveals a different side of the model's performance. The 1017 ranks second among large open-weight models on the intelligence index with a score of 45. This trails the top-scoring Claude Opus 5.5, which scored 58. The model is also among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro beats it on cost per task. These metrics suggest the model is a strong contender for high-volume workloads where speed and low cost are priorities.
Reliability remains the main concern. On the Omniscience benchmark, the model answered 43% of questions correctly. While this is among the best results for open-weight models, it lags behind proprietary systems. The model hallucinated on 67% of the questions it should have declined to answer. The benchmark's main index rewards correct answers and penalizes wrong ones, while non-answers do not count. The 1017 model scored 5 on this index, compared with 20 for Kimi K3-max and at least 42 for frontier models. This gap indicates a higher risk of generating plausible but incorrect information.
The discrepancy between efficiency and accuracy suggests the model suits specific types of work. Tasks with verifiable outputs, such as programming or OCR with human review, should work well because errors can be caught and corrected. Open-ended research or precise document interpretation is riskier. In those scenarios, hallucinations are less likely to be detected automatically. Developers using the model for such tasks may need to implement additional checks or rely on more expensive frontier models for final validation.
This trend of trading some reliability for lower costs is common among capable open-weight models. Kimi K3 launched several months ago with similar praise for its low price and effectiveness. Users reported that short fact lookups often produced invented information. Artificial Analysis data supports this observation. Despite these issues, open-weight models remain important for reducing the cost of AI agents. They also provide a check on price increases from larger commercial labs. The 1017 release continues that pattern, offering a cheaper alternative for those willing to manage its accuracy limitations.
The model is available for immediate use through DeepSeek's API and as open weights for third-party hosting. Because the weights are open, providers can run the model on their own hardware. Some third-party providers have already undercut DeepSeek's API pricing. The search feature costs $0.75 per 1,000 queries, which is below the closest rival, Parallel, at $1 per 1,000 queries. Developers can integrate the model into existing workflows without major changes to their infrastructure.
