DeepSeek 1017 Reduces Hardware Costs, Yet Accuracy Lags
The new open-weight model requires one-quarter of the previous memory but hallucinates on two-thirds of questions it should have declined.
By Gabriel Qureshi ·
DeepSeek has launched a new revision of its open-weight V4.1 Flash model, designated as 1017, which dramatically lowers memory requirements while retaining a notable tendency to generate incorrect information.
The 1017 update builds on the V4.1 Flash architecture, a Mixture-of-Experts system with 552 billion total parameters and 16 billion active parameters per run. DeepSeek positions the model for broad, open-ended tasks under the slogan "wish-driven programming made real." However, the efficiency gains come with a trade-off in reliability that persists from the previous version.
Compared to the earlier V4 Flash, the 1017 version requires only one-quarter of the High Bandwidth Memory (HBM) and one-eighth of the SSD storage for its Key-Value cache. The model now uses just 890 bytes per token entry for this cache. DeepSeek cites a cost-per-task chart based on the DeepSWE September 2 dataset, showing that both standard and 1017 variants achieve favorable results. Even low-effort runs of the 1017 model complete tasks at rates comparable to the maximum settings of more expensive competitors.
Pricing for the 1017 model at off-peak hours is set at $0.15 per million input tokens and $0.60 per million output tokens. Cache-hit input costs drop to $0.003 per million tokens. The search feature is priced at $0.75 per 1,000 queries, which is lower than the closest competitor, Parallel, at $1.00. Because the weights are open, third-party providers can host the model, often undercutting DeepSeek’s direct API rates.
Independent testing by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index with a score of 45. This trails the top-scoring proprietary model, Claude Opus 5.5, which holds a score of 58. The 1017 model is among the fastest in its category, with only Xiaomi’s MiMo V2.6-Pro offering a lower cost per task.
Reliability remains the primary drawback. On the Artificial Analysis Omniscience benchmark, the model answered 43% of questions correctly, a figure that is competitive for open-weight systems but below proprietary alternatives. The model hallucinated on 67% of questions it should have declined to answer. Its main index score is 5, compared to 20 for Kimi K3-max and at least 42 for frontier models.
The model is best suited for tasks with verifiable outputs, such as programming or format conversion, where an agent can check and retry. For open-ended research or precise document interpretation, the risk of undetected hallucinations suggests using frontier models or human review.
Source: Tom's Hardware
