DeepSeek's 1017 model uses 890 bytes per token but invents facts

DeepSeek released a new revision of its V4.1 Flash model on October 23. The update, labeled 1017, targets efficiency for open-ended tasks. The company markets the model with the slogan "wish-driven programming made real." This phrasing suggests the tool is designed for broad, descriptive instructions rather than precise commands.

The model is a Mixture-of-Experts architecture with 552 billion total parameters. Only 16 billion parameters are active during any given run. It accepts text and images and supports a context window of one million tokens. DeepSeek states that the new version requires only one-quarter of the high-bandwidth memory used by the previous V4 Flash model. It also needs one-eighth of the SSD storage for its key-value cache. Each token entry now uses just 890 bytes.

DeepSeek presents data showing the model performs between GPT-5.6 Sol and Claude Opus 5 in internal tests. It beats Anthropic's model in specific task evaluations. The cost-per-task chart built on the DeepSWE September 2 dataset places the 1017 variant in the favorable top-left region. Low-effort runs of the new model complete tasks at rates comparable to the maximum settings of more expensive competitors.

Pricing for the model is aggressive. At off-peak rates, input costs $0.15 per million tokens. Output costs $0.60 per million tokens. Cache-hit input drops to $0.003 per million tokens. The search feature costs $0.75 per 1,000 queries. This is lower than the closest rival, Parallel, which charges $1 per 1,000 queries. Because the weights are open, third-party providers can host the model and often undercut DeepSeek's direct API pricing.

Independent analysis from Artificial Analysis ranks the 1017 model second among large open-weight models. It scores 45 on the intelligence index. The top-scoring model, Claude Opus 5.5, holds a score of 58. Only Xiaomi's MiMo V2.6-Pro beats the DeepSeek model on cost per task. The 1017 model is also among the fastest in its category.

Reliability remains a major issue. On the Omniscience benchmark, the model answered 43% of questions correctly. This is among the best results for open-weight models but trails proprietary systems. The model hallucinated on 67% of questions it should have declined to answer. The benchmark penalizes wrong answers and ignores non-answers. The 1017 model scored 5, compared with 20 for Kimi K3-max and at least 42 for frontier models.

The model suits tasks with verifiable outputs, such as programming or format conversion. Agents can check work and retry if errors appear. Open-ended research and precise document interpretation are better handled by frontier models or humans. In those areas, hallucinations are less likely to be caught by automated checks.

This pattern of fast performance with loose accuracy appears common among capable open-weight models. Kimi K3 launched with similar praise for low cost and high effectiveness. However, short fact lookups often produced invented information. Open-weight models continue to serve as a check on price rises from larger labs.