Pipecat's open PhoneLLM Max Alpha 2 tops voice-agent bench at 83%, beats Sonnet 5.5
solyarisoftware · x · 2026-10-17
Pipecat Labs released PhoneLLM Alpha 2, two open-weights text models for voice agent pipelines: Mini (30B/A3.5B MoE) for speed, and Max (120B/A12B) with 'microthinking' — one sentence of reasoning that keeps multi-turn conversations accurate while staying under 200ms TTFAT. On the PhoneBench eval, Max with microthinking scored 83.0% with a 3.1% tool-call failure rate, edging out Claude Sonnet 5.5 (80.3%), while Mini (75.6%) beat Gemini 3.6 Flash. Weights are on Hugging Face; Alpha 2 also adds Pipecat Flows integration and reduced prompt-style sensitivity.
More from coding & agent
- Zach Lloyd on the 'software factory': engineering is becoming factory engineering — vikvang1 · 2026-10-17
- "A Great Model with a Bad Harness Loses": Complete Guide to Harness Engineering — aakashgupta · 2026-10-17
- Ex-Chrome lead Addy Osmani: a decent model with a great harness beats a great model — aakashgupta · 2026-10-17
- Dev uses Claude to ban millions of IPs by ASN, stopping a 0.5Tbps-scale attack — basedjensen · 2026-10-17
- GLM 5.3 Prompt for API Security Testing Sparks Debate — MicahBerkley · 2026-10-17
- RUC open-sources EvoOntology: self-evolving ontology layer lifts data agents by up to 26.7 points — udmrzn · 2026-10-17