DeepSeek's 1017 model cuts memory costs

DeepSeek released V4.1 Flash 1017 on October 23, a revision of its open-weight model that drastically lowers memory requirements. The update reduces High Bandwidth Memory usage to one-quarter of the previous version and cuts solid-state drive storage needs for the key-value cache to one-eighth. This efficiency gain means the model uses only 890 bytes per token entry, a substantial drop from prior standards.

The model features 552 billion total parameters, with 16 billion active during each run. It supports both text and image inputs and offers a context window of 1 million tokens. DeepSeek markets the release with the phrase "wish-driven programming made real," signaling a focus on open-ended tasks that rely on broad descriptions rather than precise instructions.

Pricing for the API reflects the efficiency gains. During off-peak hours, input costs stand at $0.15 per million tokens, while output costs reach $0.60 per million tokens. Cache-hit input drops to $0.003 per million tokens. The company also charges $0.75 per 1,000 queries for its search feature, a price point lower than the closest competitor, Parallel, which charges $1 per 1,000 queries.

Independent analysis by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. The model achieved a score of 45, trailing the top-scoring Claude Opus 5.5, which recorded a score of 58. In terms of speed and cost per task, the 1017 version is among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro outperforms it on cost efficiency.

Reliability remains the primary weakness. On the Omniscience benchmark, the model answered 43% of questions correctly, a rate that is high for open-weight systems but lower than proprietary alternatives. It hallucinated on 67% of questions where it should have declined to answer. The resulting index score was 5, compared to 20 for Kimi K3-max and scores of 42 or higher for frontier models.

These results suggest the model suits tasks with verifiable outputs, such as programming or format conversion, where errors can be detected and corrected. Tasks requiring precise fact lookups or open-ended research remain risky due to the high rate of invented information. The pattern mirrors earlier releases like Kimi K3, which was initially praised for cost but later found to produce inaccurate data in short queries.

The release continues the trend of open-weight models trading some reliability for lower operational costs. While frontier labs maintain higher accuracy, these efficient alternatives help keep AI agent prices stable and offer users a viable option for high-volume, low-stakes workloads.