Overview
- DeepSeek published V4.1‑Flash‑1017 on Saturday, Oct. 17 and made the build available on its website, mobile app and developer API while releasing the model weights under an MIT license on Hugging Face.
- The company’s own benchmark tables show gains such as DeepSWE v1.1 rising to 77.0 from 74.2 and Terminal‑Bench 4.0 to 42.3 from 31.2 versus the prior Flash build.
- The model uses a sparse mixture‑of‑experts design, a 1 million‑token context window and KV cache compression to cut active compute and shorten outputs, and DeepSeek publishes tiered token pricing that lowers effective per‑task bills.
- All reported benchmark scores so far come from DeepSeek’s test harnesses and independent reproductions remain pending, and customers must weigh those unverified gains against an earlier roughly five‑hour outage that raised capacity and reliability concerns.
- If third‑party tests confirm the claims, developers could shift workloads to a cheaper, self‑hostable alternative and prompt competing labs to respond on price or reasoning performance, making independent verification and production stability the key things to watch next.