890 Bytes: DeepSeek’s V4.1 Flash Cuts Memory Use by 75%

DeepSeek released a new revision of its V4.1 Flash large language model on October 23, 2026, which requires only 890 bytes per token for its key-value cache. This figure represents a 75% reduction in storage needs compared to the previous V4 Flash version. The update, labeled 1017, also cuts high-bandwidth memory requirements to one-quarter of the prior standard. These changes lower the hardware costs for running the model on long, continuous tasks.

The V4.1 Flash model is a Mixture-of-Experts architecture with 552 billion total parameters. During a run, it activates 16 billion of those parameters. The model accepts both text and image inputs and supports a context window of 1 million tokens. DeepSeek markets the update with the slogan "wish-driven programming made real," a phrase that suggests the model is designed for open-ended tasks described in broad terms.

DeepSeek’s internal benchmarks place the new model between OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Opus 5. In specific task tests, the V4.1 Flash outperforms the Anthropic model. A cost-per-task chart based on the DeepSWE September 2 dataset shows that both the standard and 1017 variants operate in the favorable top-left region. Even low-effort runs of the 1017 version complete tasks at rates comparable to the maximum settings of more expensive competitors.

At off-peak hours, DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens. Cache-hit input costs drop to $0.003 per million tokens. Because the model weights are open, third-party providers can host the software and have previously offered prices lower than DeepSeek’s direct API. The company’s search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.00.

Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It scores 45, while the top-scoring model, Claude Opus 5.5, achieves a score of 58. The V4.1 Flash is among the fastest in its category. Only Xiaomi’s MiMo V2.6-Pro offers a lower cost per task.

Reliability remains a central issue for the update. On the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly. This places it among the best open-weight models but behind proprietary systems. The model hallucinated on 67% of questions it should have declined to answer. The benchmark’s main index scores the 1017 model at 5, compared with 20 for Kimi K3-max and at least 42 for frontier models.

The source notes that this pattern of fast but loose behavior is common among capable open-weight models. When Kimi K3 launched several months ago, it was praised for being cheap and effective. However, its use for short fact lookups often produced invented information, a finding supported by the Artificial Analysis data. Open-weight models continue to serve as a way to make AI agents more affordable and to check price increases from frontier labs.

The new model is best suited for tasks with verifiable outputs. Programming, OCR with review, and format conversion work well because agents can check their work and retry if errors occur. Open-ended research, fact lookups, and precise document drafting are better handled by frontier models or humans, where hallucinations are less likely to go undetected.