Deepseek V4.1 Flash Cuts Memory Costs

Deepseek has released a revised version of its V4.1 Flash large language model, designated as 1017, which drastically lowers the hardware requirements needed for deployment. This update marks a significant shift in efficiency for the open-weight sector, offering a viable alternative to expensive proprietary systems for organizations managing large-scale AI workloads.

The architectural changes in this release focus on reducing memory overhead. The model is a Mixture-of-Experts design with 552 billion total parameters, of which 16 billion are active during any given run. It supports multimodal inputs including text and images, and features a context window of one million tokens. Compared to the previous V4 Flash iteration, the new version requires only one-quarter of the High Bandwidth Memory and one-eighth of the solid-state drive storage for its key-value cache. This reduction brings the usage down to just 890 bytes per token entry, a substantial decrease that eases the burden on local infrastructure.

Performance benchmarks suggest the model competes strongly with leading commercial offerings. In internal tests, Deepseek positions the new release between GPT-5.6 Sol and Claude Opus 5, noting superior task completion rates compared to the latter. The company’s cost-per-task analysis, based on the DeepSWE dataset from September, places both the standard and 1017 variants in the favorable region of the performance graph. Even low-effort runs of the 1017 version complete tasks at a rate comparable to the maximum settings of costlier competitors, highlighting its token efficiency.

Pricing remains a competitive advantage for the open-weight model. At off-peak rates, input costs are set at $0.15 per million tokens, with output at $0.60. Cache-hit input drops to $0.003 per million tokens. Because the weights are open, third-party providers can host the model, often undercutting Deepseek’s own API prices. The integrated search feature costs $0.75 per 1,000 queries, which is lower than the closest rival, Parallel, at $1.00.

Despite these gains, reliability remains a notable weakness. Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index with a score of 45, trailing the top-scoring Claude Opus 5.5 at 58. However, on the Omniscience benchmark, which measures the ability to know when to answer and when to decline, the model struggled. It answered 43% of questions correctly but hallucinated on 67% of questions where it should have declined. Its overall score of 5 on this benchmark compares poorly to Kimi K3-max at 20 and frontier models at 42 or higher.

These limitations suggest the model is best suited for tasks with verifiable outputs. Programming, optical character recognition with review, and format conversion are ideal applications because agents can verify results and retry if errors occur. Open-ended research, factual lookups, and the interpretation of precise documents are better handled by frontier models or human reviewers, as hallucinations are less likely to be caught automatically.

The trend of prioritizing speed and cost over strict accuracy appears to be growing among capable open-weight models. Similar issues were observed with the launch of Kimi K3 earlier this year, which was initially praised for its low cost but later found to invent information during short fact lookups. Nevertheless, open-weight models like Deepseek’s new release remain crucial for making AI agents more affordable and for providing a check on price increases from major frontier laboratories.