PhoneLLM Alpha 1: this 0.0025 $/minute open-weights voice model matches GPT-5.6 Terra — with 3x lower latency
🔎 Why a 30B model just made GPT-5.6 Terra-based voice agents obsolete
On August 28, 2026, Daily and Pipecat released PhoneLLM Alpha 1 on Hugging Face. An open-weights model with 30 billion parameters, of which only 3.5B are active at each inference. Result on PhoneBench v1: 72.3%, just one point behind GPT-5.6 Terra (72.4%), and well above Claude Sonnet 5 (68.9%).
But raw scores don't tell the whole story. Where GPT-5.6 Terra charges $0.0347 per minute with a P50 TTFAT of 980 ms, PhoneLLM shows $0.0025 and 331 ms. Thirteen times cheaper, three times faster. For a call center running 12 hours a day, the difference amounts to hundreds of thousands of dollars per month.
This is the most striking illustration to date of an underlying trend in 2026: small, open models, specialized for a specific workload, no longer just rival generalist frontier models — they surpass them on the metrics that keep a business running.
The key points
- PhoneLLM Alpha 1 is a 30B MoE model (3.5B active), full-parameter fine-tuned from NVIDIA Nemotron 3 Nano 30B-A3B, under a BSD-2 license with no commercial restrictions.
- PhoneBench v1 score: 72.3% — near parity with GPT-5.6 Terra (72.4%), behind Gemini 3.6 Flash (78.6%), ahead of Claude Sonnet 5 (68.9%).
- P50 TTFAT of 331 ms (compared to 980 ms for Terra), cost of $0.0025/minute (compared to $0.0347).
- The non-fine-tuned base model caps out at 28.6%: specialized post-training is the real performance lever.
- 80+ competing voice agents can run simultaneously on a single B200 GPU.
Recommended tools
| Tool | Main usage | Price (August 2026, check site) | Ideal for |
|---|---|---|---|
| PhoneLLM Alpha 1 | Open-weights voice agents | Free (self-host) | Call center deployment at a fixed cost |
| vLLM | Recommended serving | Free (open-source) | High-density production deployment |
| SGLang | Alternative recommended serving | Free (open-source) | Configuration flexibility |
PhoneBench v1: the numbers that matter
PhoneLLM Alpha 1 does not position itself as a generalist model. It is evaluated on PhoneBench Alpha 1, a benchmark of 15 models focused on a specific task: multi-turn tool-calling in a phone assistant context.
The ranking reveals an interesting hierarchy.
| Model | PhoneBench v1 Score | TTFAT P50 (ms) | Cost ($/min) |
|---|---|---|---|
| Gemini 3.6 Flash | 78.6% | Not published | Variable |
| GPT-5.6 Terra | 72.4% | 980 | 0.0347 |
| PhoneLLM Alpha 1 | 72.3% | 331 | 0.0025 |
| Qwen 3.8 27B | 70.0% | Not published | Variable |
| Claude Sonnet 5 | 68.9% | Not published | Variable |
Two observations stand out. First, PhoneLLM reaches 99.8% of GPT-5.6 Terra's score — a statistically negligible difference in real-world conditions. Second, Gemini 3.6 Flash dominates the benchmark, confirming that lightweight models optimized for speed have a structural advantage in voice tasks.
The telling detail: the base Nemotron 3 Nano model, without fine-tuning, plateaus at 28.6%. Specialized post-training multiplies performance by 2.5. This echoes the principle observed with CacheRL : un modèle Qwen3-4B atteint 92 % de précision en tool-calling avec 100 fois moins de compute que GPT-5. Specialization through post-training is the real multiplier of 2026, not the architecture alone.
Architecture: why 3.5B active parameters are enough for voice
PhoneLLM Alpha 1 is based on a Mixture of Experts (MoE) architecture with 30 billion total parameters but only 3.5 billion activated per token. This is a deliberate choice, not a constraint.
In voice inference, every millisecond counts. A dense 30B model would generate a prohibitive TTFAT (Time To First Audio Token). MoE makes it possible to retain the reasoning depth of a large model while limiting the actual compute to that of a small one. The result: a P95 TTFT of under 100 ms on a simple request.
The model is a full-parameter fine-tune of NVIDIA Nemotron 3 Nano 30B-A3B. No LoRA, no QLoRA — all parameters were updated during post-training. It's expensive during the training phase, but the result is a model whose weights are end-to-end optimized for a single type of interaction.
The maximum context reaches 262,000 tokens. Sufficient for long phone conversations with history, but also for injecting substantial RAG — customer files, purchase histories, medical protocols — without truncating.
This approach aligns with the philosophy of other recent open models like GLM-5.2 : le modèle open weights le plus puissant du monde — 753B MoE, 1M contexte, licence MIT, le paysage LLM bascule or MiniMax M3 : l'open-weights chinois qui défie GPT-5.5 avec 1M contexte et l'architecture MSA. The difference: PhoneLLM is not trying to be the best at everything, but the best at one specific point.
Cost and density: the calculation that changes the game
The cost analysis published by Explainx.ai corrects an initial claim by Daily (which mentioned a 94% cost saving). The real figure is 13x cheaper, not 18x. Still, that is a considerable difference.
Let's take a concrete example: a call center with 50 simultaneous active lines, 8 hours a day, 22 days a month.
With GPT-5.6 Terra at $0.0347/minute: 50 × 8 × 60 × 22 × 0.0347 = ~$18,366/month.
With PhoneLLM self-hosted on a B200 (rental at ~$3/hour from major cloud providers): the compute cost for 80+ simultaneous agents is easily absorbed. Even factoring in infrastructure, the opex drops below $2,500/month.
Density is the other key factor. 80 concurrent agents on a single B200 means that horizontal scaling remains reasonable. For 200 lines, three GPUs are enough. For an enterprise deployment, the barrier to entry in terms of hardware is minimal.
The BSD-2 license eliminates any legal uncertainty. No commercial restrictions, no burdensome attribution clause. A tech department can deploy PhoneLLM in production without a three-month legal validation.
Deployment: vLLM, SGLang, and the settings that make the difference
Daily recommends two serving engines: vLLM and SGLang. Both natively support the MoE architecture and optimize batching for real-time workloads.
Two settings are critical and counter-intuitive for anyone used to generalist models.
First, the temperature must be set to 0. In a voice agent, variability is not an asset — it's a risk. An unpredictable response on a booking tool or a medical diagnosis is a bug, not a feature. The fine-tune was designed to operate deterministically.
Next, thinking (chain-of-thought) must be disabled. Models like GPT-5.5 or Claude Opus 4.7 gain in reasoning when thinking is enabled, but at the cost of unacceptable real-time latency. PhoneLLM was trained to produce direct responses without an intermediate reasoning step. Enabling thinking degrades both latency and response quality on this specific benchmark.
For teams looking to dive deeper into the voice AI ecosystem, the full stack generally includes an STT engine upstream and a TTS engine downstream. The launch of Gemini 3.5 Transcribe by Google has shown that real-time transcription continues to progress. On the synthesis side, with ElevenLabs having crossed the 500 million dollar ARR mark, TTS quality is no longer the bottleneck. The LLM has become the central component to optimize — which is exactly what PhoneLLM addresses.
Concrete use cases: where PhoneLLM changes the game
Finance and insurance call centers
A voice agent needs to check a balance, initiate a transfer, or block a card. These actions require reliable tool-calling on internal APIs. The PhoneBench score measures exactly this. With a 72.3% multi-turn success rate, PhoneLLM is operational for the most frequent workflows — the 27.7% failure rate being mainly complex edge-case scenarios that can be escalated to a human.
Medical triage and healthcare
A 331 ms P50 latency is a game-changer for protocols where the patient needs to feel a natural interaction. The 262k context allows the entire patient record to be loaded into the system prompt. The BSD license ensures that data never leaves the facility's infrastructure — a regulatory imperative (HIPAA, health GDPR).
Retail and e-commerce
80 agents per B200 means a mid-sized retailer can deploy a 24/7 voice customer service for the cost of a single monthly salary. The return on investment is measured in weeks, not months.
Comparison with proprietary full-duplex solutions
When OpenAI launched GPT Live with full-duplex voice, the industry hailed a major breakthrough. But the proprietary model remains tied to OpenAI's infrastructure, with a cost per minute that limits massive deployments. PhoneLLM reverses the equation: the quality is nearly equivalent, but control and cost belong to the operator.
2026 Trend: Specialization beats generalization
PhoneLLM Alpha 1 is no accident. It is the symptom of a paradigm shift in AI.
In early 2025, the scaling race dominated: more parameters, more data, more compute. The agentic and generalist leaderboards were dominated by the same players — GPT-5.5 leading the pack with 98.2 in agentic and 91 in general, followed by Gemini 3.1 Pro and Claude Opus 4.7.
In 2026, the landscape fragments. Specialized open-weights models emerge in every niche: code, tool-calling, voice, long-range RAG. They are not trying to beat GPT-5.5 on the general benchmark — they beat it on the benchmark that matters to the company deploying them.
The economic logic is relentless. A frontier generalist model is expensive to serve because it has to be good at everything. A specialized 30B MoE model only mobilizes 3.5B of compute per token and excels in its domain. Full-parameter post-training, although expensive upfront, is a one-time investment that pays off with every minute of inference.
PhoneLLM is to voice what code-specialized models are to development. For those building voice agent pipelines, checking out the best AI tools for code or the best LLMs for coding shows the same dynamic: specialization wins.
Limitations: what PhoneLLM does not do
Honesty requires listing the constraints.
It's an Alpha model. The name says it all: Pipecat does not position it as production-ready for all cases. The fine-tuning was done on specific telephone interaction data. An agent that needs to handle complex contract negotiations, legal advice, or emotionally sensitive conversations is not the targeted use case.
A score of 72.3% means that more than one in four interactions fail in multi-turn tool-calling. In a call center with human escalation, this is manageable. In a fully automated workflow without a safety net, it's risky.
The model has no native multimodal capabilities. It takes text as input and produces text as output. The STT + LLM + TTS stack remains necessary, unlike natively multimodal models that are beginning to emerge.
Finally, self-hosting requires MLOps skills. vLLM and SGLang are not plug-and-play solutions. Latency monitoring, horizontal scaling, queue management — this is engineering work that proprietary APIs absorb for you.
❌ Common mistakes
Mistake 1: Enabling thinking to "improve reasoning"
This is the most frequent mistake with real-time specialized models. Thinking was designed for generalist models that need a planning step. PhoneLLM has been fine-tuned to respond directly. Enabling thinking increases latency without improving the PhoneBench score — it degrades it.
Mistake 2: Using a temperature higher than 0
In agentic voice conversation, determinism is a feature. A temperature of 0.7 or 1.0 introduces variability that results in inconsistent tool calls between two identical sessions. The Pipecat teams are clear: temperature=0.
Mistake 3: Comparing PhoneLLM to a generalist leaderboard
Seeing PhoneLLM and thinking "it's worse than GPT-5.5 (98.2 in agentic)" makes no sense. PhoneLLM was not trained for the general agentic benchmark. It was trained for PhoneBench. Comparing these scores is like evaluating a sprinter on a marathon.
Mistake 4: Deploying without an escalation safety net
A 72.3% success rate is excellent for an open-weights Alpha. It is not sufficient for total automation. Planning a transfer mechanism to a human when tool-calling fails is not optional — it is the minimum requirement for responsible deployment.
❓ Frequently Asked Questions
Can PhoneLLM Alpha 1 replace GPT-5.6 Terra in production?
Not in all cases. If your voice workflow is centered around multi-turn tool-calling with human escalation, yes — the cost/performance ratio is largely in favor of PhoneLLM. If you need general reasoning, native multimodality, or capabilities beyond a phone assistant, GPT-5.6 Terra remains more versatile.
What GPU is needed to deploy PhoneLLM?
A single NVIDIA B200 can run 80+ concurrent agents. For testing, an A100 80GB works but with reduced density. The Pipecat teams recommend vLLM or SGLang for serving.
Does the BSD-2 license allow commercial use?
Yes, without any restrictions. BSD-2 is one of the most permissive open-source licenses. You can integrate PhoneLLM into a commercial product, modify it, redistribute it — all without paying royalties or asking for permission.
Why does the base model only reach 28.6%?
Full-parameter fine-tuning is the real performance driver. The base Nemotron 3 Nano model is a good generalist model, but it was not specifically trained for phone interaction patterns. Specialized post-training injects these behaviors into the weights themselves, hence the jump from 28.6% to 72.3%.
Does PhoneLLM handle French?
The base Nemotron 3 Nano model supports multilingualism, but the PhoneLLM fine-tune was done primarily on English data. French performance has not been published and will likely be lower. This is a point to verify internally before any French-language deployment.
✅ Conclusion
PhoneLLM Alpha 1 does not seek to be the best model in the world — it seeks to be the best model for voice agents, and on this specific ground, the numbers speak for themselves: parity with GPT-5.6 Terra, 3x lower latency, 13x cheaper, 80 agents per B200. Specialized post-training is the new competitive lever for open-weights AI. To keep up with the next breakthroughs of this kind, regularly check out our latest new AI tools.