Deepseek's V4.1 Flash Slash Memory Costs While Retaining Hallucination Flaws

A newly updated open-weight model is delivering major efficiency gains while maintaining a persistent reliability issue.

Deepseek recently issued a revision labeled 1017 for its V4.1 Flash system. While it appears to be a minor patch, the changes result in a substantial reduction in resource consumption. The company markets this release with the phrase "wish-driven programming made real," signaling its utility for broad, open-ended tasks. However, independent tests show that the model still frequently generates incorrect information.

The architecture behind this release is a Mixture-of-Experts design totaling 552 billion parameters, with only 16 billion active during any given run. It supports both text and image inputs and handles a context window of one million tokens. In Deepseek's internal comparisons, the model performs between GPT-5.6 Sol and Claude Opus 5. The most notable improvement is in token efficiency, as the new version requires only one-quarter of the high-bandwidth memory and one-eighth of the solid-state storage for its cache compared to its predecessor.

Cost metrics further highlight the model's value proposition. Using DeepSWE data from September, both standard and updated variants rank in the top-left quadrant of the cost-per-task chart. Even low-effort runs of the new version match the performance of maximum settings in pricier competitors. At off-peak rates, input costs stand at $0.15 per million tokens, with output at $0.6. The search feature is priced at $0.75 per 1,000 queries, a figure lower than most rivals, with the nearest competitor charging $1.

External rankings place the updated model second among large open-weight systems on the Artificial Analysis intelligence index, scoring 45 points. This trails the top proprietary option, which recorded 58 points. The model is also among the fastest in its category, with only one other product offering a lower cost per task. Despite these strengths, the model struggles with reliability. It answered only 43% of questions correctly on the Omniscience benchmark and hallucinated on 67% of instances where it should have abstained. Its overall score of 5 falls significantly behind specialized rivals and frontier systems, which score 42 or higher.

These results suggest the model is best suited for tasks with verifiable outputs, such as coding or data conversion, where errors can be detected and corrected. It is less ideal for open-ended research or precise document interpretation, where invented facts can go unnoticed. This pattern reflects a broader trend in capable open-weight models, which often trade accuracy for speed and affordability. As frontier labs raise prices, these systems remain a critical option for budget-conscious developers and a potential check on market rates.