Deepseek V4.1 Flash Reduces Memory Use

The artificial intelligence landscape has largely divided into two distinct camps, with expensive proprietary systems competing against cheaper open-weight alternatives from Chinese developers. This dichotomy is currently shaping how companies and individuals approach the deployment of large language models for daily tasks. While frontier lab products often command high prices, they generally deliver superior consistency. In contrast, open-source rivals typically sacrifice some reliability to offer much lower operational costs, a trade-off that becomes especially apparent in long-running agentic workflows.

Deepseek has just released a new revision of its V4.1 Flash model, designated as version 1017. Although this update appears superficially minor, it delivers substantial improvements in efficiency for users. The company markets the release with the slogan, wish-driven programming made real, a phrase that signals the model's suitability for open-ended tasks defined by broad parameters rather than strict instructions. This positioning suggests a focus on the emerging trend of vibe coding, where users guide the model with high-level intent rather than detailed prompts.

Under the hood, the V4.1 Flash model utilizes a Mixture-of-Experts architecture. It contains a total of 552 billion parameters, but only 16 billion are active during any single run. The system is multimodal, accepting both text and image inputs, and supports a context window of 1 million tokens. This capacity has become a standard benchmark in the industry, ensuring the model can handle large documents and complex conversations without losing track of earlier details.

The most significant advantage of the new version lies in its resource consumption. Compared to the previous V4 Flash iteration, the 1017 update requires only one-quarter of the High Bandwidth Memory and one-eighth of the solid-state drive storage for its key-value cache. Each token entry now consumes just 890 bytes. Deepseek supports these claims with cost-per-task data derived from the DeepSWE September 2 dataset. Both variants of the new model appear in the favorable top-left region of the chart, indicating high performance at low cost. Even low-effort runs match the completion rates of the maximum settings used by more expensive competitors.

Pricing for the new model remains competitive, especially when considering off-peak rates. Input costs are set at $0.15 per million tokens, while output costs $0.60. Cache-hit inputs drop to just $0.003 per million. Because the weights are open, third-party providers can host the model and often undercut the official API pricing. The integrated search feature costs $0.75 per 1,000 queries, a rate lower than most rivals, with Parallel being the closest competitor at $1.00.

Independent evaluations place the 1017 model as the second highest ranked among large open-weight systems on the Artificial Analysis intelligence index. It achieves a score of 45, whereas the top-scoring Claude Opus 5.5 reaches 58. The model is also among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro offering a lower cost per task. However, these impressive metrics are tempered by notable weaknesses in reliability.

On the Omniscience benchmark, the model answered only 43% of questions correctly. While this is a strong result for an open-weight system, it lags behind proprietary alternatives. More concerning, the model hallucinated on 67% of the questions it should have declined to answer. Since the benchmark penalizes wrong answers and ignores non-responses, the 1017 model scores a 5. This compares poorly to Kimi K3-max, which scores 20, and frontier models that typically score 42 or higher.

These results suggest that the new Deepseek model is best suited for tasks with verifiable outputs. Programming, optical character recognition with human review, and format conversion are good fits because an agent can verify the work and retry if errors occur. For open-ended research, precise fact lookups, or drafting critical documents, users should rely on frontier models or human oversight to catch potential hallucinations. This pattern of capable but loose behavior is becoming common among open-weight models, as seen with the earlier Kimi K3 release. Despite these reliability issues, open-weight models remain vital for keeping AI agents affordable and providing a check on the pricing power of major commercial labs.