Deepseek V4.1 Flash cuts memory use

Deepseek released a new revision of its open-weight V4.1 Flash model on October 23, 2026. Labeled 1017, the update significantly lowers the hardware requirements for running the system while maintaining high processing speeds.

The model operates as a Mixture-of-Experts architecture with 552 billion total parameters. During any given run, only 16 billion of these parameters are active. It supports text and image inputs and handles context windows of up to 1 million tokens. In its own benchmarks, Deepseek places the new version between GPT-5.6 Sol and Claude Opus 5, noting it outperforms the Anthropic model in specific task tests.

The most notable change is in storage efficiency. Compared with the previous V4 Flash version, the 1017 update requires only one-quarter of the high-bandwidth memory and one-eighth of the solid-state drive storage for its KV cache. This reduces the memory footprint to 890 bytes per token entry. Deepseek provided cost-per-task data from the DeepSWE September 2 dataset to support these figures. The data shows that even low-effort runs of the new model complete tasks at rates similar to the maximum settings of more expensive competitors.

Pricing for the service remains low. At off-peak rates, Deepseek charges $0.15 per million input tokens and $0.6 per million output tokens. If the system uses a cache hit for input, the cost drops to $0.003 per million tokens. The company’s search feature costs $0.75 per 1,000 queries, which is lower than its closest rival, Parallel, at $1 per 1,000 queries.

Independent testing by Artificial Analysis ranks the 1017 model second among large open-weight models on its intelligence index. It achieved a score of 45, trailing the top-scoring Claude Opus 5.5, which earned 58. The model is also among the fastest in its category, with only Xiaomi’s MiMo V2.6-Pro offering a lower cost per task.

Reliability remains a primary concern for users. On the Artificial Analysis Omniscience benchmark, the model answered only 43% of questions correctly. It hallucinated on 67% of questions where it should have declined to answer. The benchmark’s main index scored the 1017 model at 5, compared with 20 for Kimi K3-max and at least 42 for frontier proprietary models.

These results suggest the model is best suited for tasks with verifiable outputs, such as programming or format conversion, where an agent can check and retry work. Tasks requiring precise fact-lookup or document interpretation may still require a frontier model or human review to catch errors. The pattern mirrors the launch of Kimi K3 earlier this year, which was praised for cost but criticized for inventing information during short lookups.

Deepseek describes the update with the slogan "wish-driven programming made real," a phrase that references vibe coding and suggests the model fits open-ended tasks. The open-weight nature of the model allows third-party providers to host it, often at prices lower than Deepseek’s own API.