Skip to content
CallMissed
Book a demo

Explore CallMissed

Back to all models
LLM Chatnewfasttools

phoneLLM mini

by Pipecat · Released 2026

Pipecat's early-access open-weights model for voice pipelines where speed matters most. 32K context, tool calling, text only.

LLM Chat

phoneLLM mini

Powered by Pipecat · Mixture-of-experts (Nemotron 3 Nano base)

Context Window

32K

Parameters

30B MoE (3.5B active)

Max Output

Not disclosed

Category

LLM Chat

Overview

phoneLLM mini is an open-weights model from Pipecat, based on NVIDIA Nemotron 3 Nano: a 30B mixture-of-experts model with 3.5B active parameters, trained on synthetic voice-agent sessions. It is meant for voice pipelines where speed matters most.

On CallMissed you call it with the id `deepgram-phonellm-mini` on the OpenAI-compatible `/v1/chat/completions` endpoint, or pick it as the language model of a voice agent (speech-to-text, language model, text-to-speech). It has a 32,768-token context window, supports tool calling and streaming, and takes text only. Voice agents use temperature 0.0 by default, which is the recommended setting.

It does not reason: any `reasoning_effort` is sent as `none`, because the model would otherwise put its reasoning in the spoken answer. Measured from our India servers, a one-sentence reply takes 0.3 to 0.4 seconds. Hindi is adequate but terse; Tamil is weak.

It is excluded from the provider's model training. It is processed outside India and is not available to India-only keys or voice agents hosted in India. As an early-access model, request fields and limits may change and there is no SLA.

Pricing

MetricPrice
Input /1M tokens₹35.0000
Output /1M tokens₹90.0000

1 credit = ₹1 ≈ US$0.0104 (US$1 = ₹96). Transparent per-call pricing — you pay only for what you call, with no per-seat charge or minimums.

Key Highlights

  • 0.3 to 0.4s for a one-sentence reply from our India servers
  • Open weights, based on NVIDIA Nemotron 3 Nano
  • Excluded from the provider's model training
  • 32,768-token context window

Benchmarks

BenchmarkScore
Reply latency0.3-0.4s
Active parameters3.5B
Context32K

Technical Details

  • 30B mixture-of-experts, 3.5B active parameters, based on NVIDIA Nemotron 3 Nano
  • Context window: 32,768 tokens; text input and output only
  • Tool calling with function tools; tool_choice "auto" or "none" only
  • Streaming supported; response_format (JSON mode) is not supported
  • No reasoning: any reasoning_effort is sent as none
  • Recommended temperature 0.0; max_tokens, top_p, stop, seed and penalties are ignored
  • Early access: fields and limits may change; no SLA

Strengths

  • Among the fastest replies in our voice tests
  • Open weights, excluded from provider training
  • Tool calling for agent workflows

Limitations

  • Early access: request fields and limits may change, no SLA
  • 32K context and text only (no images or files)
  • Processed outside India: refused for India-only keys and hosting "india" agents
  • No reasoning and no JSON mode; Hindi is terse, Tamil is weak; paid plans only

Use Cases

Voice agentsPhone receptionistsLatency-critical voice pipelines

API Example

curl https://api.callmissed.com/v1/chat/completions \
  -H "Authorization: Bearer cm_YOUR_KEY" \
  -d '{"model": "deepgram-phonellm-mini", "temperature": 0, "messages": [{"role": "user", "content": "Confirm my appointment for Friday"}]}'

Endpoint: POST /v1/chat/completions · Model ID: deepgram-phonellm-mini

Try phoneLLM mini now

Get 1000 free API credits on signup. No credit card required.

Start free Book a DemoRead docs