Overview
- DeepSeek rolled out V4.1-Flash-1017 on October 17 and made the model live on its website, app and developer platform while publishing the weights under an MIT license on Hugging Face.
- The company reports higher benchmark scores, saying the DeepSWE v1.1 coding score rose to 77.0% and that the model produces shorter outputs and finishes tasks faster.
- V4.1-Flash-1017 uses a sparse mixture-of-experts design with about 552 billion parameters and a 1 million token context window, and DeepSeek highlights KV-cache compression and FP4 quantization to lower inference cost.
- DeepSeek opened Search & Fetch developer APIs with a temporary free tier through November 1 and published public pricing that emphasizes cache-hit and shorter-output discounts to reduce effective per-task fees.
- Independent verification and production stability remain open questions after a roughly five-hour outage earlier in the week, and wider testing will determine whether the company’s cost and performance claims hold in real-world use.