Inception previews Mercury Voice, targeting sub-170ms latency

The diffusion-LLM maker wants voice agents fast enough for the phone; you can build on Mercury 2.5 today, plus fresh gateway models and an Ollama Mac speedup.

Nowline SEP 30 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Mercury Voice: a preview aimed at picking up the phone

    Inception previewed Mercury Voice on Sep 29, targeting under 170ms to first token — roughly the line where a voice agent stops sounding like a bot waiting to answer. It extends the company's diffusion approach, which drafts tokens in parallel instead of one at a time. It's a preview with no GA date yet, so watch the waitlist rather than rewiring your stack today.

  • You can ship the voice stack on Mercury 2.5 now

    No need to wait for Voice: Mercury 2.5 already runs at 1,000+ tokens/sec with sub-300ms reasoning traces, priced at $0.25/$1.00 per Mtok — on the order of a cent per call-minute. It's OpenAI-compatible and drops into ElevenLabs, Vapi, LiveKit, Pipecat and Retell pipelines, so an ASR→LLM→TTS voice agent is a config change, not a rebuild.

  • Fireworks' Ember-1 lands on Vercel AI Gateway

    Ember-1, a research-preview model tuned for coding and agent workflows, is now callable through Vercel AI Gateway (Sep 27). Its pitch is lower reasoning-token usage, so your agent loops get cheaper without touching your SDK — swap the model string and A/B it against your current default.

  • A free stealth model, 'Pixel Canary,' on AI Gateway

    Vercel quietly opened Pixel Canary for free on AI Gateway (Sep 25), billed as strong at coding, frontend, and mobile app design. Free stealth access is a weekend-build unlock: point a frontend or app-scaffolding project at it and see how it stacks up before it gets a real name and a price.

  • Ollama flips Apple Silicon to MLX by default

    In v0.40.0-rc0 (Sep 25), Ollama runs MLX-supported model architectures on the MLX runtime by default on Apple Silicon — faster local inference with no flags to set. If you run models on a Mac, grab the release candidate and expect a speed bump for free.