Full write-up: wiring Nemotron brain, Phonon voice, and WeVisOne-VL eyes into a local agent
lmoroney · x · 2026-10-27
Laurence Moroney's full post assembles three open releases into a desk-scale cloud-free agent:
- Brain — Nemotron 3.5 Bolt E4B (NVIDIA, Oct 23): post-trained from Gemma-4-E4B-it for local voice/device assistants; 8B params (4B effective), OpenMDW-1.1 license, 4GB at Q4KM; conversational/thinking toggles, multi-turn tool use, 64K context on NVIDIA's card; claims 40% higher average scores vs Nemotron 3 Nano 4B and up to 80% fewer output tokens on agent suites; already in Ollama v0.40.10 — though Ollama lists it as a VLM with 128K context, conflicting with NVIDIA's 64K, so verify locally; 15.5 tok/s on Jetson Orin Nano 8GB.
- Voice — Phonon (Ollama 0.40.11, Oct 24): speech models on macOS starting with Gradium's Phonon; ollama run phonon plus an OpenAI-compatible /v1/audio/speech endpoint so existing TTS code just needs a new base URL.
- Eyes — WeVisOne-VL: grounds camera crops with boxes and masks.
He wires them as specialists with no cloud API key and warns that vendor packaging claims deserve hands-on verification.
Related event: Three Open-Source Models Combine Into a Desktop-Ready Local Agent(2 posts)→
More from coding & agent
- 'Codex decided': devs admit making up reasons in code review for agent-written code — FlatStill2540 · 2026-10-27
- Running AI agents on a legacy IBM WMS: enforce tool order, script checks, verify evidence — ksrijith · 2026-10-27
- 365 public MCP servers tested: 51% lack tool input descriptions, the most common bug — abhiramreddykalluru · 2026-10-27
- CoreWeave ARIA agent ships four new live in-chat panels for ML run analysis — _ScottCondron · 2026-10-27
- opentel-mcp brings OpenTelemetry silent-failure detection to Python MCP servers — Thirumalaiboobathi · 2026-10-27
- A hands-on local-model benchmark: porting an old screensaver to Python OpenGL with Qwen 3.8 — prusswan · 2026-10-27