AI News Roundup: DeepSeek Search APIs, Reflection Beam Weights, and PhoneLLM Alpha 2

Here is the AI news from the last few days that matters for teams building agents and voice workflows. Every item links to its source.

DeepSeek updates its Flash model and opens Search and Fetch APIs

On October 17, DeepSeek began serving DeepSeek-V4.1-Flash-1017 under the existing `deepseek-flash` model name, so developers don't need to change their calls. DeepSeek also documented two standalone endpoints for agents: Search, which returns ranked web results, and Fetch, which returns a page as Markdown. The English pricing lists each at $0.75 per 1,000 successful requests, with failed or empty requests not billed. (DeepSeek API Docs, Ahead of the Bot)

Reflection releases Beam's open weights

On October 16, Reflection made Beam available as open weights under the Apache 2.0 license. It is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion active per token, and a 1M-token context window, aimed at coding and agentic work. It ships in BF16, FP8, and NVFP4 formats with day-one support in transformers, vLLM, and SGLang. (Reflection, Hugging Face)

Pipecat releases PhoneLLM Alpha 2 for voice agents

Earlier this week, Pipecat Labs released PhoneLLM Alpha 2, two open-weight text models designed to sit inside voice agent pipelines. PhoneLLM Mini is a 30B mixture-of-experts model tuned for speed, and PhoneLLM Max is a 120B model for stronger accuracy. The update adds Pipecat Flows integration and less sensitivity to prompting style. (Pipecat Labs)

Cheaper models and faster voice pipelines make it easier for a small business to answer every lead the moment it comes in. If you want an AI voice agent that qualifies callers and hands ready buyers to a human closer, see Ozkra or get in touch.