Deepseek V4.1 Flash Reduces Memory Footprint While Retaining Hallucination Issues

Deepseek has released a new revision of its open-weight V4.1 Flash model, designated as version 1017. While the update appears superficially minor, it delivers a substantial improvement in operational efficiency. The company markets the release with the phrase wish-driven programming made real, a slogan that suggests the tool is well suited for broad, open-ended tasks often associated with vibe coding. However, independent evaluations indicate that the model remains prone to generating incorrect information.

The large language model market is currently divided into two distinct camps. Closed commercial models from leading research labs offer high performance but come with significant operating costs. In contrast, open-weight models from Chinese competitors typically sacrifice some reliability to achieve much lower running expenses, a trade-off that is particularly beneficial for long-running agentic workflows.

V4.1 Flash operates as a Mixture-of-Experts architecture with 552 billion total parameters, activating only 16 billion during any given run. The system is multimodal, processing both text and images, and supports a context window of one million tokens, a standard now common in the industry. According to Deepseek's internal benchmarks, the model performs between GPT-5.6 Sol and Claude Opus 5, even surpassing Anthropic's model in specific task tests. The most notable improvements are in performance per task and token efficiency.

Compared to the previous V4 Flash version, the new 1017 model requires only one-quarter of the high-bandwidth memory and one-eighth of the SSD storage for its key-value cache. This efficiency is achieved through a compact structure of 890 bytes per token entry. Deepseek supports these claims with a cost-per-task chart based on the DeepSWE September 2 dataset. Both the standard and 1017 variants appear in the favorable top-left region of the chart, with low-effort runs of the 1017 version completing tasks at rates comparable to the maximum settings of more expensive competitors.

Pricing for the model is competitive, especially during off-peak hours. Input costs are set at $0.15 per million tokens, while output costs are $0.60 per million tokens. Cache-hit input drops to $0.003 per million tokens. Because the model weights are open, it can be hosted by third-party providers who have previously offered lower rates than Deepseek's own API. Additionally, the built-in search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.00 per 1,000 queries.

On the Artificial Analysis intelligence index, the 1017 model in maximum mode ranks second among large open-weight models with a score of 45. This compares to a score of 58 for the top-scoring Claude Opus 5.5 in maximum mode with fallback. The model is among the fastest in its category, and only Xiaomi's MiMo V2.6-Pro offers a lower cost per task.

The primary drawback of the new model is its reliability. In the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly, a strong result for open-weight systems but below proprietary alternatives. It hallucinated on 67% of questions where it should have declined to answer. The benchmark's main index rewards correct answers and penalizes wrong ones, while non-answers receive no credit. The 1017 model scored 5 on this index, compared to 20 for Kimi K3-max and at least 42 for frontier models.

These results suggest the model is capable but best suited for jobs with verifiable outputs. Tasks such as programming, optical character recognition with review, and format conversion work well because agents can check and retry their work. Open-ended research, fact lookups, and precise document interpretation are better left to frontier models or humans, as hallucinations are less likely to be caught in these areas. This fast-and-loose behavior appears to be a trend among capable open-weight models. When Kimi K3 launched several months ago, it was praised for being cheap and effective, but its use in short fact lookups often produced invented information. Despite these issues, open-weight models remain important for making AI agents more affordable and as a check on price increases from frontier labs.