DeepSeek V4.1 Flash 1017 reduces memory usage by 75 percent

DeepSeek has released a revised version of its open-weight V4.1 Flash model, designated as version 1017. This update changes how much memory a server needs to run the software. The revision reduces the memory footprint for the KV cache by 75 percent compared to the previous V4 Flash model. This efficiency gain allows the system to operate on less expensive hardware.

The model uses a Mixture-of-Experts architecture with 552 billion total parameters. Only 16 billion parameters are active during any single run. It supports text and image inputs and handles a context window of one million tokens. DeepSeek describes the release with the slogan 'wish-driven programming made real,' a reference to coding based on broad instructions rather than precise code.

In its own performance tests, DeepSeek places the 1017 model between GPT-5.6 Sol and Claude Opus 5. The company states the new version requires one-quarter of the high-bandwidth memory and one-eighth of the SSD storage used by the older V4 Flash model. Specifically, the system uses just 890 bytes per token entry for its cache. A cost-per-task chart based on the DeepSWE September 2 dataset shows the 1017 variant completing tasks at a rate similar to the maximum settings of more expensive competitors.

DeepSeek lists the off-peak input price at $0.15 per million tokens. The output price is $0.60 per million tokens, and cache-hit input drops to $0.003. Because the model weights are open, third-party providers can host the software and have previously offered lower API rates. The company charges $0.75 for 1,000 search queries, a rate lower than the closest rival, Parallel, which charges $1.

Independent testing by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It scores 45 on that index, compared to 58 for the top-scoring Claude Opus 5.5. The model is among the fastest in its category. Only Xiaomi’s MiMo V2.6-Pro beats it on cost per task.

Reliability remains a concern for the new release. On the Omniscience benchmark, the model answered 43 percent of questions correctly. It is among the best open-weight models in that category but trails proprietary systems. The model hallucinated on 67 percent of questions it should have declined to answer. Its main index score is 5, whereas Kimi K3-max scores 20 and frontier models score at least 42.

These results suggest the model fits specific workflows best. Tasks with verifiable outputs, such as programming or format conversion, work well because agents can check and retry their work. Open-ended research or precise document drafting carries a higher risk because errors are harder to detect. The trend of fast but loose behavior appears common among capable open-weight models.

Kimi K3, launched several months ago, was praised for being cheap and effective. However, short fact lookups sometimes produced invented information, a pattern supported by Artificial Analysis data. Open-weight models remain important for keeping AI agents affordable and providing a check on price increases from frontier labs.