THE CRUNCH
DeepSeek has updated its V4.1 Flash open-weight model to revision 1017, and the revision is a bigger deal than the version number suggests. The model is a Mixture-of-Experts design with 552B total parameters, of which 16B are active during runs. It accepts text and images, handles a 1 million token context window, and DeepSeek markets it for open-ended tasks with the slogan "wish-driven programming made real", a nod to vibe coding.
The efficiency gains are the headline numbers. Compared with V4 Flash, the new revision uses one quarter of the HBM and one eighth of the SSD storage for its KV cache, at 890 bytes per token entry. On DeepSeek's own pricing, off-peak input tokens cost $0.15 per million and output $0.6, dropping to $0.003 per million on cache hits. Third-party providers have undercut DeepSeek's API before, so those figures could fall further.
On benchmarks, DeepSeek's internal tests place the model between GPT-5.6 Sol and Claude Opus 5, and on a cost-per-task graph based on DeepSWE's dataset even low-effort runs of DSF-1017 land near the task completion rates of pricier competitors. Artificial Analysis ranks it second among large open-weight models on its intelligence index with a score of 45, against 58 for Claude Opus 5.5, while noting it is one of the fastest in its class.
The catch is reliability. On Artificial Analysis' Omniscience benchmark, the model answered correctly 43% of the time, strong for open-weight models but behind proprietary rivals, and it hallucinated 67% of its non-answers. Its main index score of 5 compares with 20 for Kimi K3-max and 42 or more for frontier models. That makes it well suited to tasks with verifiable outputs, such as programming, OCR with review and format conversions, where an agent can check its own work, but a poor choice for open-ended research or drafting precise documents.