Release notes
Changelog.
What changed, and when.
Loom multi-session conductor + Tensor Code preview
- new
Loom, our autonomous coding agent, now drives multiple Claude Code sessions in parallel from one brief — with a live plan you can edit before it executes. Available at loom.mingllm.com; docs at docs.mingllm.com.
- new
Tensor Code — the CLI coding agent — opens in preview. Repo-wide context, diff-based edits, and a test-runner loop that re-runs until checks pass.
- fix
Tensor's DOM observer no longer declares a virtualized page 'settled' early on infinite-scroll feeds; citations land on the right span more often.
Rocky answer-engine surface + cross-surface receipts
- new
Rocky joins the lineup — a personal answer-engine surface that runs the same local base model and writes to the same Orb memory and Receipts log as Jarvis and Tensor.
- change
The Receipts log is now shared across every surface. An action taken in Tensor is visible (and undoable, where safe) from Jarvis and Rocky, not just the app that took it.
- fix
Fixed memory mis-attribution when two facts shared a name; the temporal re-ranking head now ships by default.
Receipts log + Orb consolidation overhaul
- new
Every Tensor and Jarvis action now appears in a unified Receipts log with a timestamp, inputs, and a one-tap undo where applicable.
- change
Orb memory consolidation runs every 90 seconds of idle time instead of on a fixed schedule. Multi-session recall improves measurably on long days.
- fix
Jarvis no longer trips its own wake word when reading output aloud.
Tensor side-panel citations
- new
Every Tensor answer now pins citations back to the exact DOM range on the source tab. Click to jump; the span highlights on arrival.
- fix
Fixed a race where Tensor synthesised answers before slow-loading tabs finished reading.
saturday-4b ships as the default base
- new
MingLLM now runs on saturday-4b, our unified 4B base tuned via 11 rounds of expert iteration. End-to-end browser task score: 94.9% on our internal 59-task set.
- change
Inference latency reduced ~22% on M2/M3 Max. Time-to-first-token sits under 1s for most voice interactions.
Local First by default
- change
Cloud calls are now off by default. MingLLM runs entirely on-device unless you explicitly grant a remote tool permission.
- new
Added a privacy panel that enumerates every outgoing request from the last 30 days.
First public preview
- new
MingLLM preview opens on invite. Jarvis + Tensor + Orb memory.