Deepseek's 1017 revision cuts V4.1 Flash KV cache storage needs to one-eighth

Deepseek has shipped a revision of its open-weight V4.1 Flash model, labeled 1017, that slashes the storage required for its KV cache. Compared with the prior V4 Flash, the update needs one-quarter the HBM and one-eighth the SSD storage for that cache, using 890 bytes per token entry.

V4.1 Flash is a Mixture-of-Experts model carrying 552 billion total parameters, with 16 billion active during a run. It accepts text and images and supports a context window of 1 million tokens. Deepseek's own benchmarks place it between GPT-5.6 Sol and Claude Opus 5, and ahead of Anthropic's model on task tests.

Deepseek published a cost-per-task chart built on DeepSWE's September 2 dataset. Both the standard and 1017 variants fall in the favorable top-left region of that chart, and low-effort runs of the 1017 version complete tasks at a rate comparable to the max settings of costlier competitors. At Deepseek's off-peak rates, input runs $0.15 per million tokens, output $0.6, and cache-hit input $0.003. The company's search feature costs $0.75 per 1,000 queries; the closest listed rival, Parallel, charges $1.

Artificial Analysis ranks the 1017 model at max settings second among large open-weight models on its intelligence index, scoring 45 against 58 for Claude Opus 5.5 at max with fallback. The service also rates it among the fastest in its category, with only Xiaomi's MiMo V2.6-Pro beating it on cost per task.

Reliability remains the weak point. On Artificial Analysis' Omniscience benchmark, the model answered 43% of questions correctly, among the best for open-weight models but behind proprietary systems. It hallucinated on 67% of questions it should have declined to answer. The benchmark's main index penalizes wrong answers, and the 1017 model scores 5 there, against 20 for Kimi K3-max and 42 or more for frontier models.

Deepseek's product page carries the slogan "wish-driven programming made real." Tasks with verifiable outputs, including programming, OCR with review and format conversion, can be checked and retried by an agent. Open-ended research, fact lookups, and drafting or interpreting precise documents remain better suited to a frontier model or a human, since hallucinations in those tasks go undetected.