DeepSeek's V4.1 Flash 1017 slashes KV cache to 890 bytes per token
The update reduces hardware demands by 75% but retains a high rate of hallucinations on fact-based queries, according to independent benchmarks.
By Harper Laurent ·
DeepSeek released version 1017 of its open-weight V4.1 Flash model on October 23, 2026. The update reduces the hardware required to run the model while maintaining its core capabilities. The company describes the release with the slogan "wish-driven programming made real," suggesting the tool is designed for broad, open-ended tasks often associated with vibe coding.
The revision significantly lowers storage requirements. The new model uses 890 bytes per token for its KV cache. This figure represents a 75% reduction compared to the previous V4 Flash version. Specifically, the update requires one-quarter of the High Bandwidth Memory (HBM) and one-eighth of the Solid State Drive (SSD) storage for the same cache size.
V4.1 Flash is a Mixture-of-Experts model with 552 billion total parameters. Only 16 billion of these parameters are active during any given run. The model supports both text and image inputs and handles a context window of 1 million tokens. These specifications align with current standards for large language models in the open-weight sector.
DeepSeek positions the model competitively against major commercial offerings. In internal benchmarks, V4.1 Flash performs between GPT-5.6 Sol and Claude Opus 5. It surpasses the Anthropic model in specific task tests. The company also presents cost-per-task data based on the DeepSWE September 2 dataset. This data shows that low-effort runs of the 1017 version match the completion rates of maximum-setting configurations from costlier competitors.
Pricing for the model remains low. DeepSeek lists input costs at $0.15 per million tokens and output costs at $0.60. Cache-hit input costs drop to $0.003 per million tokens. The search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.00. Because the weights are open, third-party providers can host the model, often undercutting DeepSeek's direct API prices.
Independent testing by Artificial Analysis places the 1017 model second among large open-weight models on its intelligence index. The model scored 45 on the index. This trails the top-scoring Claude Opus 5.5, which achieved a score of 58. The model is among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro offering a lower cost per task.
Reliability remains a concern for the update. On the Omniscience benchmark, the model answered only 43% of questions correctly. While this is strong for open-weight models, it lags behind proprietary systems. The model hallucinated on 67% of questions it should have declined to answer. On the benchmark's main index, which penalizes wrong answers but not non-answers, the 1017 model scored 5. This compares to 20 for Kimi K3-max and 42 or more for frontier models.
The performance profile suggests the model suits tasks with verifiable outputs. Programming, OCR with human review, and format conversion allow agents to check and retry results. Open-ended research and precise document interpretation carry higher risks because errors are harder to detect. The pattern of high efficiency paired with frequent hallucinations mirrors trends seen in recent open-weight releases, such as Kimi K3.
Open-weight models continue to serve two primary roles. They lower the cost of running AI agents and provide a check on pricing from frontier labs. Despite the hallucination issues, the reduced hardware demands of V4.1 Flash 1017 make it a practical option for workloads where verification is built into the workflow.
