DeepSeek's new model uses 890 bytes of memory per token

DeepSeek released a new revision of its V4.1 Flash open-weight model on October 23.

The update, designated 1017, reduces the hardware required to run the model by a substantial margin. The company describes the release with the slogan "wish-driven programming made real," a nod to the practice of coding via broad, natural language prompts rather than specific code.

The underlying architecture remains a Mixture-of-Experts model with 552 billion total parameters. During any given run, only 16 billion of those parameters are active. The system handles text and images and supports a context window of 1 million tokens, which has become a standard feature in the sector.

Efficiency is the primary change in this revision. Compared to the previous V4 Flash version, the 1017 model requires one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its key-value cache. Each token entry now consumes just 890 bytes. This reduction directly lowers the cost of running the model on independent hardware.

DeepSeek provided a cost-per-task chart based on the DeepSWE September 2 dataset. The data places both the standard and 1017 variants in the top-left region of the graph, indicating high performance at low cost. Low-effort runs of the 1017 version complete tasks at a rate similar to the maximum settings of more expensive competitors.

Pricing for the API remains low. At off-peak rates, input costs $0.15 per million tokens and output costs $0.60. When a cache hit occurs, the input cost drops to $0.003 per million tokens. The company's search feature costs $0.75 per 1,000 queries, a rate lower than its closest rival, Parallel, which charges $1 per 1,000 queries.

Independent evaluations place the model as second among large open-weight systems on the Artificial Analysis intelligence index. The 1017 model achieved a score of 45. This trails the top-scoring proprietary model, Claude Opus 5.5, which scored 58. Only Xiaomi's MiMo V2.6-Pro offers a lower cost per task in this category.

Reliability remains a weak point for the update. On the Omniscience benchmark, the model answered 43% of questions correctly. While this is among the best results for open-weight models, it lags behind proprietary systems. More notably, the model hallucinated on 67% of questions where it should have declined to answer. Its main index score was 5, compared to 20 for Kimi K3-max and 42 or higher for frontier models.

Experts suggest the model is best suited for tasks with verifiable outputs, such as programming or format conversion, where an agent can check its work. Open-ended research or precise document interpretation carries a higher risk of undetected errors.

The pattern of trading accuracy for affordability is consistent with recent releases from Chinese open-weight labs. This approach allows developers to build more affordable AI agents while providing a price check on the proprietary frontier.