New Deepseek Model Cuts Hardware Costs by 75% but Struggles with Accuracy

Deepseek has released a new revision of its V4.1 Flash model, designated as version 1017, which drastically reduces the hardware requirements needed to run the system. While the update appears minor on the surface, it delivers a substantial efficiency gain for developers and businesses looking to deploy large language models without massive infrastructure overheads. The company describes the update with the slogan "wish-driven programming made real," suggesting it is designed for open-ended tasks that are described in broad, conversational terms rather than strict technical specifications.

The technical improvements are stark. The model is a Mixture-of-Experts architecture with 552 billion total parameters, though only 16 billion are active during any single run. It supports text and image inputs and handles a context window of one million tokens. Compared to the previous V4 Flash version, the new 1017 update requires only one-quarter of the High Bandwidth Memory and one-eighth of the SSD storage for its key-value cache. Specifically, it uses just 890 bytes per token entry, a significant drop that makes long-running agentic work much cheaper to execute.

Several key figures define the model's current standing in the market. Deepseek claims performance sits between GPT-5.6 Sol and Claude Opus 5, with task test results even exceeding Anthropic's model. Independent data from Artificial Analysis places 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 which holds a score of 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.

The pricing structure reflects this efficiency. At off-peak rates, input tokens cost $0.15 per million and output tokens cost $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 direct API prices. Deepseek's search feature is priced at $0.75 per 1,000 queries, which is cheaper than the closest rival, Parallel, at $1.00.

The people involved in this shift include developers who can now self-host these massive models due to the reduced storage needs. However, the primary stakeholder is the enterprise sector looking to balance cost with reliability. The model is currently best suited for tasks with verifiable outputs, such as programming or format conversion, where an agent can check its work and retry if errors occur. It is less ideal for open-ended research or fact lookups, where errors are harder to detect.

The main drawback remains reliability. On the Omniscience benchmark, the model answered only 43% of questions correctly. More concerning, it hallucinated on 67% of the questions it should have declined to answer. While it scores a 5 on the benchmark's main index, this is behind Kimi K3-max at 20 and frontier models at 42 or more. This pattern of fast but loose behavior is becoming a trend among capable open-weight models, which remain important for keeping AI affordable and checking price hikes from larger labs.

Despite the accuracy issues, the efficiency gains are significant. The model allows for cheaper, faster deployment of AI agents. For many use cases, the trade-off between lower cost and higher hallucination rates is acceptable, especially when the output can be programmatically verified. This positions the 1017 update as a strong contender for high-volume, low-complexity tasks in the current AI landscape.