DeepSeek V4.1 Flash 1017 Cuts Memory Use to 890 Bytes per Token

DeepSeek released a new revision of its open-weight V4.1 Flash model, designated 1017, on Wednesday. The update focuses on efficiency, reducing the memory required to run the model. The company describes the release with the slogan "wish-driven programming made real," a reference to vibe coding that positions the tool for open-ended tasks described in broad terms.

The model is a Mixture-of-Experts system with 552 billion total parameters. During a run, 16 billion parameters remain active. It accepts text and images and supports a context window of 1 million tokens. According to DeepSeek, the 1017 version requires only one-quarter of the High Bandwidth Memory (HBM) and one-eighth of the Solid State Drive (SSD) storage for its key-value cache compared to the previous V4 Flash iteration.

The new architecture uses just 890 bytes per token entry. DeepSeek provided a cost-per-task chart based on the DeepSWE September 2 dataset to back these claims. In that analysis, both the standard and 1017 variants appear in the favorable top-left region. Low-effort runs of the 1017 version complete tasks at a rate comparable to the maximum settings of costlier competitors.

Pricing for the new model remains low. At off-peak hours, input costs 15 cents per million tokens, and output costs 60 cents. Cache-hit input drops to 0.3 cents. The search feature costs 75 cents per 1,000 queries, which is lower than rivals, with Parallel being the closest competitor at 1 dollar.

Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It scores 45, trailing the top-scoring Claude Opus 5.5, which scores 58. 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 Artificial Analysis' Omniscience benchmark, the 1017 model answered 43% of questions correctly. This result places it among the best open-weight models but behind proprietary systems. The model hallucinated on 67% of questions where it should have declined to answer. The benchmark index scores the 1017 model at 5, compared to 20 for Kimi K3-max and at least 42 for frontier models.

The model suits tasks with verifiable outputs. Programming, OCR with review, and format conversion work well because an agent can check its work and retry. Open-ended research, fact lookups, and drafting precise documents are better left to frontier models or humans, as hallucinations are less likely to be caught in those scenarios.

Open-weight models like this continue to serve as a check on price rises from frontier labs. They make AI agents more affordable, though the fast-and-loose behavior appears to be a trend among capable open-weight systems. Kimi K3 launched several months ago with similar praise for cost and effectiveness, but users reported invented information in short fact lookups.