DeepSeek V4.1 Flash slashes memory costs but struggles with accuracy

DeepSeek has introduced a revised version of its V4.1 Flash model, designated as 1017, creating a distinct tension in the artificial intelligence market. While the update promises dramatic reductions in hardware requirements, independent assessments suggest the model still suffers from significant reliability issues. This dichotomy highlights a broader divide in the sector, where closed commercial models from major laboratories command premium prices for high performance, while Chinese open-weight alternatives often sacrifice some accuracy for lower operational costs.

The technical specifications of the new release point to a major efficiency jump. The system operates as a Mixture-of-Experts architecture with a total of 552 billion parameters, though only 16 billion are active during any given run. It supports both text and image inputs and handles a context window of one million tokens, a standard that has become common in recent years. The primary selling point is the reduction in memory footprint. Compared to the previous V4 Flash iteration, the 1017 update requires just one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its cache. This results in a token entry size of only 890 bytes, a figure that significantly lowers the barrier to entry for enterprise deployment.

DeepSeek supports these claims with cost-per-task data derived from the DeepSWE benchmark, which was updated in September 2026. In these tests, both the standard and 1017 variants appear in the favorable top-left region of the cost-performance chart. The data indicates that even low-effort runs of the new model complete tasks at a rate comparable to the maximum settings of more expensive competitors. While DeepSWE is just one of many available metrics, the results suggest a compelling value proposition for businesses looking to optimize their compute spend.

Pricing structures further enhance the model's appeal for budget-conscious developers. At DeepSeek's off-peak rates, input costs stand at $0.15 per million tokens, while output is priced at $0.60 per million tokens. If the system utilizes cache hits, the input cost drops to $0.003 per million tokens. Because the weights are open, third-party providers have previously managed to undercut DeepSeek's own API pricing. Additionally, the company's search feature is charged at $0.75 per 1,000 queries, a rate that is lower than its nearest rival, Parallel, which charges $1.00.

Performance rankings place the new model competitively within its class. Artificial Analysis ranks the 1017 version second among large open-weight models on its intelligence index, assigning it a score of 45. This trails the top-scoring Claude Opus 5.5, which achieved a score of 58 with fallback settings enabled. In terms of speed, the model is among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro offering a better cost-per-task ratio. DeepSeek's own internal benchmarks position the model between GPT-5.6 Sol and Claude Opus 5, with the latter being outperformed in specific task tests.

However, the efficiency gains come with a notable drawback in reliability. On the Omniscience benchmark from Artificial Analysis, the model answered only 43% of questions correctly. While this is among the best results for open-weight systems, it lags behind proprietary alternatives. More critically, the model hallucinated on 67% of the questions it should have declined to answer. The benchmark's primary index rewards correct responses and penalizes errors, while non-answers are not counted. Under this metric, the 1017 model scored a 5, compared to a 20 for Kimi K3-max and scores of 42 or higher for frontier models.

These reliability metrics suggest the model is best suited for specific workflows. Tasks with verifiable outputs, such as programming, optical character recognition with human review, and format conversion, are likely to perform well. In these cases, an agent can check its work and retry if errors occur. Conversely, open-ended research, precise fact lookups, and the drafting of detailed documents are better left to frontier models or human reviewers. In these scenarios, hallucinations are less likely to be caught automatically, posing a greater risk to data integrity.

The pattern of high capability paired with loose accuracy appears to be a growing trend among open-weight models. When Kimi K3 launched several months ago, it was praised for its low cost and effectiveness. However, user experience often ended quickly after short fact lookups produced invented information, a finding supported by Artificial Analysis data. Despite these issues, open-weight models remain a vital component of the AI ecosystem. They help make autonomous agents more affordable and serve as a potential check on price increases from major frontier laboratories.

For investors and enterprise buyers, the release underscores a shifting dynamic in the market. The ability to run powerful models on less expensive hardware is a significant economic advantage. Yet, the persistence of hallucinations means that the choice between open and closed models cannot be made on cost alone. The decision will increasingly depend on the specific nature of the task and the tolerance for error in the output.