Roadmap
What's next.
We publish the roadmap so the tradeoffs are legible. Shipping dates are honest estimates, not promises; scope is the variable.
Jarvis voice + Tensor browser public preview
Live now: an invite-gated public preview of Jarvis (voice) and Tensor (browser) with Orb memory on macOS. Invites are going out in waves; the next batch opens through the Q3 build.
Receipts log across all surfaces
Shipped. Every action MingLLM takes — across voice, browser, code, and the Rocky answer engine — lands in one unified log. Undo where safe, review where it matters.
Loom autonomous coding agent
Live at loom.mingllm.com — an autonomous coding agent that conducts multiple sessions in parallel from a single brief, with a plan you edit before it executes. Docs at docs.mingllm.com.
Tensor Code general availability
The CLI coding agent graduated from internal use to preview in May; GA brings repo-wide context, diff-based edits, and test-runner integration to everyone.
8GB memory mode
Current builds target 16GB unified memory. Q3 lowers the floor so MingLLM runs on 8GB M1 / M2 Airs at reduced but usable quality.
Gemma-4 27B MoE base
Upgrade the base model from 4B dense to a ~4B-active-parameter MoE. Same latency, meaningfully stronger reasoning.
Windows + Linux
We built for macOS first to keep scope tight. Windows and Linux builds are queued after the v1 release is stable.
Shared memory (household + team)
The current Orb is strictly single-user. Multi-principal memory — shared between family members or small teams — is a 2027 research project.