Daily Releases PhoneLLM Alpha 2 With Two Open-Weights Voice Agent Models

Two models for different priorities

PhoneLLM Mini keeps the 30B/A3.5B MoE architecture of Alpha 1. Daily recommends it when speed and size matter most. PhoneLLM Max is a 120B/A12B model, aimed at cases that need stronger intelligence and better accuracy while still delivering what the company calls industry-leading inference speed.

Besides accuracy gains from an expanded dataset, Daily lists three main changes in Alpha 2: microthinking in PhoneLLM Max, support for Pipecat Flows, and reduced sensitivity to different prompting styles.

Microthinking and Pipecat Flows

Most reasoning LLMs produce paragraphs or even pages of thought per request, which raises cost and latency. Daily says that letting PhoneLLM Max output a single sentence of reasoning has a measurable effect on whether it calls the correct tools. The company describes this as a compromise between full reasoning and having reasoning disabled, and says one sentence at PhoneLLM's inference speed has a negligible effect on response time.

Pipecat Flows is a framework for structured conversations with voice agents. A programmatic harness guides the model through a complex sequence of steps while keeping each step simple, which Daily argues suits smaller, cheaper models. Alpha 2 was trained on synthetic Pipecat Flows sessions, so it already understands the format and can be used as a drop-in replacement for larger models, according to the company.

In tests with real customers, Daily says PhoneLLM Max brings calls to a successful resolution at a higher rate than GPT-4.1 when integrated into Flows, with around half the latency and a fraction of the cost.

The training dataset now covers more use cases and prompting styles, which Daily says should mean less time tweaking prompts.

Updated PhoneBench benchmark

Daily also updated PhoneBench to cover cases the previous model struggled with. It now checks whether the model can correct earlier mistakes in a conversation, follow sequences of instructions and handle Pipecat Flows scenarios. The benchmark reports separately how often the model errs when calling a tool and how often it makes a major mistake of any kind during a conversation.

Where and how to run it

Daily's partners AWS, Baseten, Cloudflare and Modal offer PhoneLLM as click-to-deploy models and hosted API endpoints. On Cloudflare's Workers AI platform, PhoneLLM Mini is priced at $0.06 per M input tokens and $0.24 per M output tokens. Modal offers optimized endpoints for both Mini and Max.

Weights are available for self-hosting in several formats:

• PhoneLLM Mini NVFP4, which fits on an NVIDIA RTX 5090
• PhoneLLM Max NVFP4, which fits on an NVIDIA DGX Spark
• PhoneLLM Mini FP8, for NVIDIA H100
• PhoneLLM Mini BF16 and PhoneLLM Max BF16, recommended for production serving at scale

For vLLM, Daily advises setting temperature to 0.0, turning reasoning off for Mini and on for Max to enable microthinking. On a B200 GPU, the company measured Mini at 17ms time to first token and aggregate decode of 10k tok/s. Max was tested on 2x B200 and should perform similarly on 2x B300, with TTFT of about 27ms and aggregate decode of about 6k tok/s. Max also supports MTP, which Daily says can double decode speed when few parallel calls are active. Max requires thinking to be on, and Daily is considering training future versions to work with it either on or off.

How it was trained

Mini is based on NVIDIA's Nemotron 3 Nano, with 30B total and 3.5B active parameters. Max is built on Nemotron Super 3.5 VL, a 120B model with 12B active parameters. Training uses NVIDIA's NeMo Framework, including Megatron Bridge and NeMo RL, and Max needs 16x B300 GPUs across two machines.

The data is fully synthetic and designed to mimic real-world voice agent use cases. Open-weights teacher models generate responses that the student model learns to imitate, across varied prompting styles, tool sets and harnesses such as Pipecat Flows. Releases so far are full-weights fine-tunes using SFT. Daily has seen improvements from a final GRPO stage and says it will likely enable it in future releases once it is confident in the quality.

What it means for developers

For teams building voice agents, Alpha 2 offers open-weights options at two sizes, with hosted endpoints from several providers and self-hosting on hardware ranging from a single RTX 5090 to B200 clusters. The comparison with GPT-4.1 and the speed figures come from Daily's own testing, so results in other setups may differ.