DeepSeek 1017 model drops memory footprint to 890 bytes per token

Developers running large language models on their own hardware will see a sharp drop in storage requirements as DeepSeek rolls out the 1017 revision of its V4.1 Flash model. The update reduces the key-value cache size to 890 bytes per token. This change cuts the high-bandwidth memory requirement by 75 percent and reduces solid-state drive storage needs to one-eighth of the previous version.

DeepSeek markets the model with the slogan "wish-driven programming made real." The company positions the tool for open-ended tasks that users describe in broad terms. The model features 552 billion total parameters, with 16 billion active during a run. It supports text and image inputs and handles a context window of 1 million tokens. In internal benchmarks, DeepSeek places the model between GPT-5.6 Sol and Claude Opus 5. The company claims the 1017 variant outperforms Anthropic's model in specific task tests.

Cost remains a central feature of the release. At off-peak rates, DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens. Cache-hit input drops to $0.003 per million tokens. The search feature costs $0.75 per 1,000 queries. Third-party providers have previously offered lower prices than DeepSeek's API because the model weights are open. A cost-per-task chart based on the DeepSWE September 2 dataset shows the 1017 variant competing effectively with more expensive alternatives. Low-effort runs of the new model match the performance of maximum-effort settings from costlier rivals.

Independent assessments reveal a trade-off between efficiency and reliability. Artificial Analysis ranks the 1017 model second among large open-weight systems on its intelligence index. It scores 45, compared to 58 for the top-scoring Claude Opus 5.5. The model ranks among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro offers a lower cost per task. However, the model answers only 43% of questions correctly on the Omniscience benchmark. It hallucinates on 67% of questions where it should have declined to answer. The main index score is 5, while Kimi K3-max scores 20 and frontier models score 42 or higher.

The performance gap suggests the model suits tasks with verifiable outputs. Programming, optical character recognition with review, and format conversion allow agents to check their work and retry errors. Open-ended research and precise document drafting remain risky because hallucinations are less likely to be caught. This pattern mirrors previous releases like Kimi K3, which was praised for low cost but criticized for inventing information during fact lookups. The 1017 update continues the trend of capable open-weight models that prioritize speed and affordability over strict factual accuracy.

The release underscores the division in the market between closed commercial models and Chinese open-weight rivals. Frontier labs offer high reliability at high cost, while open-weight models trade some accuracy for lower running expenses. This distinction is particularly relevant for long-running agentic work. The 1017 model provides a cheaper processing path for developers who can verify outputs. It remains a strong option for specific workflows despite its limitations in general fact retrieval.