Overview
- DeepSeek pushed the V4.1‑Flash‑1017 update live on its website, app and API on Oct. 17, and posted the model weights and code to Hugging Face under the MIT license.
- The company says the point release is a Pareto improvement that shortens outputs and runtimes while raising benchmark scores, citing a DeepSWE v1.1 coding score of 77.0 versus 74.2 for the prior build.
- Technically, DeepSeek reports a Mixture‑of‑Experts design with a 552 billion‑parameter backbone, a 1 million‑token context window, KV‑cache compression and reinforcement learning with self‑feedback to boost agentic and coding workflows.
- DeepSeek published low API pricing and cache policies that greatly reduce per‑task bills and opened Search & Fetch APIs (with promotional free/discount windows), and Arena.ai placed the model as the top open model on a WebDev leaderboard.
- All performance claims so far come from DeepSeek’s own tests and early leaderboard placements, and a roughly five‑hour outage earlier the same week highlights unresolved infrastructure and independent‑verification risks that could limit enterprise adoption.