LM Studio 1.1.3 brings private on-device voice transcription to Linux
Dictate to a local model with audio that never leaves your machine; the app also adds inline agent images. Elsewhere: Edge0 runs a 35B model in 3GB.
Copy markdown
Speech to text, none of it leaves your box
LM Studio 1.1.3 (Bionic) adds real-time voice transcription on Linux with NVIDIA GPUs, turning mic audio into prompt text without any cloud service ever touching your recordings. It was macOS/Windows-only before this.
The catch: chat window only, no audio API yet
Transcription lives in the desktop chat UI; the release notes ship no OpenAI-compatible audio endpoint, so you can't wire dictation into automated pipelines yet. AMD GPU acceleration is still in progress.
Also in 1.1.3: inline agent images and org MCPs
Agent-created images now render inline in conversations, llama.cpp jumps to 2.38.0, and enterprises get organization-managed MCP servers — plus a fix for the Linux symlink loops that were breaking file watching.
Elsewhere: Edge0 streams a 35B model in 3GB on a Mac
Edge0's Apache-2.0 preview runs a 35B MoE (256 experts, 4 active) by streaming experts off SSD, peaking near 2.9GB active memory at 15-18 tok/s on an M4 Pro — a 35B-class model, fully offline, on a laptop. Apple Silicon/MLX today, CUDA on the roadmap.