The report says V4.1 uses about one-quarter of the HBM and one-eighth of the SSD KV-cache storage of V4 Flash, with a KV-cache entry of 890 bytes per token.
The article also cites DeepSeek’s cost-per-task chart, where low-effort 1017 runs sit near proprietary models run at maximum effort. It presents coding and other verifiable work as a more suitable use case than open-ended factual research.
For reliability, Tom’s Hardware cites Artificial Analysis’s Omniscience benchmark: DSF-1017 at maximum effort answered correctly 43% of the time and hallucinated on 67% of non-answers. It scored 5 on the benchmark’s main index, versus 20 for Kimi K3-max and 42 or more for frontier models. The cited Intelligence Index was 45, versus 58 for Claude Opus 5.5 max with fallback.
The source does not establish why these hallucinations occur, nor does it establish performance across all deployments, prompts, or agent workflows.