DeepSeek V4.1 Flash 1017 reduces KV cache to 890 bytes per token
The new open-weight model cuts hardware costs by 75 percent but hallucinates on 67 percent of questions it should have declined.
By Lucy Lefebvre ·
DeepSeek released V4.1 Flash 1017, a revision of its open-weight large language model that drastically lowers memory requirements. The update reduces the high-bandwidth memory needed for the key-value cache to one-quarter of the previous V4 Flash version. SSD storage needs drop to one-eighth, with each token entry consuming just 890 bytes. This efficiency allows the model to process long-running agentic tasks at a lower cost than many competitors.
The 1017 build is a Mixture-of-Experts architecture with 552 billion total parameters, of which 16 billion are active during a run. It accepts text and images and supports a context window of 1 million tokens. DeepSeek describes the release with the slogan "wish-driven programming made real," targeting users who provide broad task descriptions. The company’s internal benchmarks place the model between GPT-5.6 Sol and Claude Opus 5 in overall performance. In specific task tests, it outperforms Anthropic’s model.
Pricing for the API remains low. Off-peak input costs are $0.15 per million tokens, while output costs are $0.60. If the model hits the cache, input costs drop to $0.003 per million tokens. DeepSeek’s search feature charges $0.75 per 1,000 queries, which is cheaper than the closest rival, Parallel, at $1.00. Because the weights are open, third-party providers can host the model and have previously offered rates below DeepSeek’s own API.
Independent testing by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It scored 45 on the index, trailing the top-scoring Claude Opus 5.5, which earned 58. The model is among the fastest in its category, and only Xiaomi’s MiMo V2.6-Pro beats it on cost per task. DeepSeek’s cost-per-task chart, based on the DeepSWE September 2 dataset, shows the 1017 variant landing in the top-left region, indicating high performance at low cost. Low-effort runs of the new model match the output of "max" settings on more expensive competitors.
Reliability remains a weak point. On the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly. It hallucinated on 67% of questions it should have declined to answer. The benchmark’s main index rewards correct answers and penalizes errors, while non-answers do not affect the score. The 1017 model scored 5 on this index, compared to 20 for Kimi K3-max and 42 or more for frontier proprietary models.
These results suggest the model fits specific workflows. Tasks with verifiable outputs, such as programming, OCR with human review, and format conversion, work well because agents can check their work and retry. Open-ended research, fact lookups, and drafting precise documents are better suited to frontier models or human review, where hallucinations are less likely to go unnoticed. The pattern mirrors the launch of Kimi K3 several months ago, which was praised for cost but criticized for inventing information during short fact lookups.
The release positions DeepSeek in a market split between expensive, high-accuracy closed models and cheaper, less reliable open-weight rivals. The 1017 model offers a clear efficiency gain for developers who can tolerate or verify errors. It provides a lower-cost option for running AI agents, acting as a potential check on price increases from major frontier labs.
Source: Tom's Hardware
