Deepseek V4.1 Flash Cuts Memory Costs but Struggles with Accuracy

Deepseek has rolled out a revised version of its open-weight V4.1 Flash model, designated as build 1017. While the label suggests a minor point release, the update appears to deliver substantial efficiency improvements for developers and businesses seeking lower operating costs. The company markets the release with the slogan "wish-driven programming made real," a phrase that hints at the growing trend of vibe coding, where users describe broad, open-ended tasks rather than specific technical commands. However, independent assessments suggest that this efficiency comes with a notable trade-off in reliability.

The current large language model market is generally divided into two camps. Frontier laboratories offer closed commercial models that are expensive but highly accurate. In contrast, Chinese open-weight competitors typically sacrifice some reliability to provide significantly lower running costs, a feature that is particularly valuable for long-running agentic workflows where continuous operation is required.

Technically, V4.1 Flash is a Mixture-of-Experts architecture with a total of 552 billion parameters. Only 16 billion of these parameters are active during any given run, which helps manage computational load. The model is multimodal, accepting both text and image inputs, and supports a context window of one million tokens, a standard that has become common in the industry.

Deepseek’s internal benchmarks position the model between GPT-5.6 Sol and Claude Opus 5, with the new version even outperforming Anthropic’s model in specific task tests. The most significant improvements, however, are found in token efficiency and hardware requirements. Compared to the previous V4 Flash, the 1017 version requires only one-quarter of the High Bandwidth Memory and one-eighth of the SSD storage for its KV cache. It achieves this with a memory footprint of just 890 bytes per token entry.

To substantiate these claims, Deepseek provided a cost-per-task chart based on the DeepSWE dataset from September 2026. Both the standard and 1017 variants appear in the favorable upper-left region of the graph, indicating high performance at low cost. Notably, even low-effort runs of the 1017 version complete tasks at a rate comparable to the maximum settings of more expensive competitors. While DeepSWE is just one of many available benchmarks, the results are striking for the open-weight sector.

Pricing remains competitive, especially for those running the model on their own infrastructure due to its open weights. Deepseek’s off-peak API pricing lists input tokens at $0.15 per million and output tokens at $0.60, with cache-hit input dropping to $0.003 per million. The company’s search feature costs $0.75 per 1,000 queries, which is lower than its closest rival, Parallel, at $1.00. Third-party providers have previously undercut these API rates, further lowering the barrier to entry.

Independent data from Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index, with a score of 45. This trails the top-scoring Claude Opus 5.5, which holds a score of 58. The model is among the fastest in its class, with only Xiaomi’s MiMo V2.6-Pro offering a lower cost per task. However, the model faces criticism regarding its reliability in broader contexts.

The primary drawback is the model’s tendency to hallucinate. On the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly, a strong result for open-weight systems but a step behind proprietary leaders. More concerning, it hallucinated on 67% of questions where it should have declined to answer. On the benchmark’s main index, which penalizes wrong answers, the 1017 model scored 5, compared to 20 for Kimi K3-max and at least 42 for frontier models. This suggests the model is best suited for tasks with verifiable outputs, such as coding or OCR with human review, rather than open-ended research where errors might go undetected.

This pattern of fast and loose behavior appears to be a recurring trend among capable open-weight models. Previous releases, such as Kimi K3, were initially praised for their affordability but later revealed to be prone to inventing information during short fact lookups. Despite these reliability issues, open-weight models remain a critical component of the AI ecosystem, offering a check on rising prices from frontier labs and making AI agents more accessible to working developers.