DeepSeek releases V4.1 Flash 1017 with reduced memory requirements
The new open-weight model uses 890 bytes per token for KV cache storage, though independent tests show it still struggles with factual accuracy.
By Ivan Dvorak ·
Chinese artificial intelligence firm DeepSeek launched a revised version of its V4.1 Flash model on Thursday. The update, designated 1017, focuses on improving hardware efficiency for running the system. The company describes the revision as a minor point update, but the changes alter how the model consumes memory during operation.
DeepSeek stated that the V4.1 Flash 1017 requires only one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its key-value cache compared to the previous version. This reduction results in a usage rate of 890 bytes per token entry. The model is a Mixture-of-Experts architecture with 552 billion total parameters, of which 16 billion are active during inference. It supports text and image inputs and processes context windows of up to one million tokens.
In DeepSeek's internal benchmarks, the new model performs between OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5. The company reported that the 1017 variant beats Anthropic's model in specific task tests. A cost-per-task chart based on the DeepSWE September 2 dataset shows that low-effort runs of the 1017 model complete tasks at rates similar to the maximum settings of more expensive competitors. DeepSeek's off-peak pricing sets input costs at $0.15 per million tokens and output costs at $0.60 per million tokens.
Independent testing by Artificial Analysis provides a different perspective on the model's capabilities. On its intelligence index, the 1017 model in max mode ranks second among large open-weight models with a score of 45. This trails the top-scoring Claude Opus 5.5, which received a score of 58. The 1017 model is among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro offering a lower cost per task.
Reliability remains a concern for the new release. On the Omniscience benchmark, the model answered 43% of questions correctly, a rate that is strong for open-weight models but lower than proprietary systems. The model hallucinated on 67% of questions where it should have declined to answer. Its score on the benchmark's main index was 5, compared to 20 for Kimi K3-max and at least 42 for frontier models.
These results suggest the model is best suited for tasks with verifiable outputs, such as programming or format conversion. Agents can check work and retry in these scenarios. Open-ended research or precise document drafting may require frontier models or human review because hallucinations are harder to detect. The trend of fast but loose behavior appears common among capable open-weight models, a pattern also noted with the earlier Kimi K3 launch.
The release maintains DeepSeek's position in the open-weight market, where lower running costs are traded for some reliability. Third-party providers can run the open weights and have previously undercut DeepSeek's API prices. The search feature costs $0.75 per 1,000 queries, which is lower than rivals such as Parallel at $1.00 per 1,000 queries.
