Phonely opens Alma, a voice-native LLM built for low-latency phone agents

Phonely has released Alma, a language model designed specifically for real-time voice agents rather than general chat, coding or long-horizon agent work. The model is trained around phone-call behavior and is intended to handle interruptions, spoken identifiers, tool calls and tightly structured conversational flows with lower latency than general frontier models.\n\nThe specialization matters because voice-agent failures are often different from text-chat failures. A model can be strong at reasoning but still feel unusable on a call if first-token latency is high, if it mishandles interruptions or if it skips a required verification or transfer step. Phonely says Alma was trained and evaluated on millions of real AI phone conversations and exposes both hosted and self-hostable deployment options.\n\nIndependent contact-center publication No Jitter confirmed the launch on September 4 and interviewed Phonely CEO Will Bodewes about the decision to optimize specifically for voice. The product is not positioned as a replacement for a general-purpose frontier model: Phonely’s own site explicitly says Alma is not built for creative writing, coding or long-horizon multi-agent workflows and recommends pairing it with a broader model when those capabilities are needed.\n\nPhonely publishes unusually detailed performance claims. Its current Alma page reports sub-185ms average response latency, about 206ms at the slowest 1% of measured turns, 97% function-call accuracy and materially lower blended token pricing than the general models in its test. It also advertises self-hosting for sensitive calls, zero training on customer data, configurable retention, PII redaction, SOC 2 Type II, HIPAA/BAA availability and GDPR support.\n\nThose numbers should be treated as provider benchmarks, not a neutral leaderboard. Phonely says the comparisons used 200 held-out call transcripts with identical prompts and grading; that is useful methodology disclosure, but AiToolMap has not independently reproduced the latency, accuracy, cost or carbon claims. The stronger evidence today is that Alma has a concrete product surface, explicit scope limits and independent launch coverage.\n\nFor the directory, Alma is a credible new model/tool candidate because it represents a growing pattern in agent infrastructure: specialized models optimized for one operational constraint rather than general intelligence. Canonical review should verify API pricing, self-hosting terms, security attestations and whether third-party voice-agent teams can actually reproduce the claimed latency and reliability advantages.