DeepSeek 1017 reduces cache memory to 890 bytes per token

DeepSeek has released revision 1017 of its V4.1 Flash model, reducing the memory required for each token entry to 890 bytes. This figure represents a 75 percent reduction compared to the previous V4 Flash version. The company states that the model now requires only one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its key-value cache.

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 handles a context window of 1 million tokens. DeepSeek markets the release with the slogan "wish-driven programming made real," positioning it for open-ended tasks defined by broad descriptions rather than strict specifications.

Pricing for the new model reflects its efficiency. DeepSeek lists input costs at $0.15 per million tokens and output costs at $0.60. Cache-hit inputs cost $0.003 per million tokens. Its search tool charges $0.75 per 1,000 queries, a rate below its closest competitor, Parallel, which charges $1.00. Because the weights are open, third-party providers can host the model, potentially driving costs lower than DeepSeek's own API rates.

Benchmarks show mixed results for the 1017 revision. Artificial Analysis ranks it 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 task efficiency tests, the model performs comparably to the high-effort settings of costlier competitors. Xiaomi's MiMo V2.6-Pro remains the only model in this category that undercuts DeepSeek on cost per task.

Reliability remains a weak point. On the Artificial Analysis Omniscience benchmark, the model answered 43 percent of questions correctly. It hallucinated on 67 percent of questions where it should have declined to answer. The benchmark scores the model at 5, compared to 20 for Kimi K3-max and 42 or more for frontier models. The scoring system rewards correct answers, penalizes incorrect ones, and assigns no value to non-answers.

These results suggest the model suits tasks with verifiable outputs, such as programming or format conversion, where agents can check and retry work. Open-ended research or precise document drafting remains difficult due to the high rate of invented information. The model sits in a market split between expensive, reliable closed models and cheaper, less reliable open-weight alternatives. Its lower memory footprint lowers hardware requirements, making it a viable option for cost-sensitive deployments despite accuracy gaps.

The release follows a trend among capable open-weight models that prioritize speed and cost over strict factual accuracy. Kimi K3 faced similar criticism months after its launch for producing invented information in short fact-lookup tasks. Despite these issues, open-weight models continue to influence pricing in the AI market by offering a lower-cost alternative to frontier lab offerings.