890 bytes: How Deepseek's V4.1 Flash slashes memory demands

For organizations running large language models on their own hardware, the cost of storage has just dropped significantly. Deepseek has released a new revision of its V4.1 Flash model, designated as version 1017. While the label suggests a minor update, the changes offer a substantial efficiency gain. 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 compared to the previous V4 Flash. Each token entry now uses just 890 bytes, which allows for cheaper long-running operations and makes the model more accessible for independent developers and smaller enterprises.

The release of version 1017 follows a pattern in the large language model market that is increasingly divided into two camps. Closed commercial models from major frontier labs tend to be expensive but highly reliable. In contrast, open-weight models from Chinese developers often trade some reliability for much lower running costs. This trade-off is particularly notable for long-running agentic work. Deepseek's new model fits squarely into this second category, aiming to make advanced AI capabilities more affordable while accepting a higher risk of error.

The V4.1 Flash architecture is a Mixture-of-Experts system with 552 billion total parameters. However, only 16 billion parameters are active during any given run, which contributes to its efficiency. The model is multimodal, meaning it can process both text and images. It also supports a context window of 1 million tokens, a standard that has become common in the industry. Deepseek markets the update with the slogan "wish-driven programming made real," a nod to the trend of vibe coding where users describe tasks in broad terms rather than precise code.

Performance benchmarks suggest the model is competitive with top-tier proprietary systems. In Deepseek's own testing, V4.1 Flash performs between GPT-5.6 Sol and Claude Opus 5. It even beats Anthropic's model in certain task tests. The efficiency gains are the standout feature. A cost-per-task chart built on the DeepSWE September 2 dataset shows that both the standard and 1017 variants land in the favorable top-left region. Low-effort runs of the 1017 version complete tasks at a rate comparable to the maximum settings of more expensive competitors.

Pricing remains a major selling point for the model. At Deepseek's off-peak rates, input costs $0.15 per million tokens, while output costs $0.60. If the model hits the cache, the input cost drops to $0.003 per million tokens. Because the weights are open, third-party providers can host the model and have previously undercut Deepseek's API pricing. The company's search feature is also priced competitively at $0.75 per 1,000 queries, which is lower than rivals like Parallel, which charges $1.00.

Independent evaluations from Artificial Analysis place the 1017 model second among large open-weight models on its intelligence index. It scores 45 on the index, compared to 58 for the top-scoring Claude Opus 5.5. The model is also among the fastest in its category. Only Xiaomi's MiMo V2.6-Pro beats it on cost per task. These figures suggest that the model offers a strong balance of speed, intelligence, and cost for those who can tolerate its occasional inaccuracies.

The primary drawback is reliability. On the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly. While this is among the best results for open-weight models, it trails proprietary systems. The model hallucinated on 67% of questions it should have declined to answer. The benchmark's main index rewards correct answers and penalizes wrong ones, while non-answers do not affect the score. The 1017 model scores 5 on this index, compared to 20 for Kimi K3-max and 42 or more for frontier models.

This reliability gap means the model is best suited for specific types of work. Tasks with verifiable outputs, such as programming, optical character recognition with review, and format conversion, work well because an agent can check its work and retry if needed. However, open-ended research, fact lookups, and drafting precise documents are better left to frontier models or humans. This is because hallucinations in these areas are unlikely to be caught automatically.

The trend of fast and loose behavior appears to be common among capable open-weight models. When Kimi K3 launched several months ago, it was praised for being cheap and effective. However, user experiences quickly turned negative when short fact lookups produced invented information. This issue is supported by Artificial Analysis data. Despite these challenges, open-weight models remain important for making AI agents more affordable and serve as a potential check on price increases from frontier labs.

For now, the 1017 version of V4.1 Flash offers a compelling option for developers who prioritize cost and speed over perfect accuracy. The dramatic reduction in memory requirements makes it easier to deploy in constrained environments. As the market continues to split, users must weigh the lower costs of open-weight models against the higher reliability of proprietary systems. The choice will depend on the specific needs of the application and the ability to verify outputs.