DeepSeek V4.1 Flash 1017 Uses 890 Bytes Per Token
The updated model reduces hardware demands by 75% but continues to hallucinate on two-thirds of questions it should decline.
By Joseph Jenkins ·
DeepSeek’s latest revision of its V4.1 Flash model, labeled 1017, requires only 890 bytes of cache storage per token. This figure represents a 75% reduction in High Bandwidth Memory usage compared to the previous V4 Flash version. The update also cuts Solid State Drive storage requirements to one-eighth of the original amount. These efficiency gains lower the hardware barrier for running the 552 billion parameter Mixture-of-Experts model, which activates 16 billion parameters during processing.
The model accepts text and images and supports a 1 million token context window. DeepSeek describes the update as enabling wish-driven programming, a nod to open-ended coding tasks. In the company’s internal benchmarks, V4.1 Flash performs between OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Opus 5. It surpasses the Anthropic model in specific task tests. A cost-per-task chart based on DeepSWE’s September 2 dataset shows that low-effort runs of the 1017 variant match the completion rates of maximum-effort settings in more expensive competing models.
Pricing remains low for open-weight options. DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours. Cache-hit input costs drop to $0.003 per million tokens. The search feature costs $0.75 per 1,000 queries, undercutting rival Parallel, which charges $1.00. Because the weights are open, third-party providers can host the model, often at lower rates than DeepSeek’s own API.
Reliability remains the model’s primary weakness. On the Artificial Analysis Omniscience benchmark, the 1017 model answered 43% of questions correctly. While this rank is among the best for open-weight models, it trails proprietary systems. The model hallucinated on 67% of questions where it should have declined to answer. It received a score of 5 on the benchmark’s main index, compared to 20 for Kimi K3-max and at least 42 for frontier models.
Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index with a score of 45. The top-scoring Claude Opus 5.5 achieved 58. Only Xiaomi’s MiMo V2.6-Pro beats V4.1 Flash on cost per task. The model’s speed places it in the upper tier of its category.
The efficiency gains make the model suitable for tasks with verifiable outputs, such as programming or optical character recognition. Agents can check and retry these tasks to catch errors. Open-ended research and precise document interpretation remain better suited to frontier models or human review, as hallucinations are harder to detect in those contexts. The release continues a trend where capable open-weight models trade some reliability for lower costs.
DeepSeek’s update arrives as the market splits between expensive closed commercial models and cheaper Chinese open-weight rivals. The 1017 revision offers a cheaper path for running agentic workloads, though users must accept a higher rate of invented information on factual queries.
Source: Tom's Hardware
