Roadmap
Where Local Quench has been and where it is going. Plans, not promises: the order changes when tests or users show something else matters more, and a harness change lands only when a benchmark shows it helps on real local models.
- Released
- Built, in testing
- Planned
- 0.1.0
First public beta
Released October 2026
- A coding agent built around small local models, with a harness that catches what they get wrong.
- QuenchBench: a coding benchmark that runs on your own graphics card.
- Windows, Linux, and an Android app that drives the computer from a phone.
- 0.1.1
Security and privacy
Released 27 October 2026
- Six rounds of security review and their fixes, a read-only sandbox, signed release manifests.
- The first shared leaderboard results.
- 0.1.2we are here
Protect, voice and your models as helpers
Built, in testing; planned for November
- Quench Protect preview: scanners, an AI audit that hunts for flaws, fixes on a branch of their own that you review and merge.
- Your local models as helpers for Claude Code, Codex and Claude Desktop.
- Voice input on your computer, the Permissions Center, chat mode, a terminal, the Model Compatibility Doctor.
- Tablets in desktop mode; 151 security tasks in QuenchBench Protect.
- Everything in 0.1.2
- 0.1.3 – 0.1.4
Apple Silicon and the details
Planned
- macOS on Apple Silicon with MLX models: a beta for testers first. Want to test it?
- The experimental harness behaviours measured on QuenchBench and turned on where they help.
- Better diagnostics when a model or an engine misbehaves.
- 0.1.5
Quench Protect, complete
Planned
- Out of preview: the AI audit measured against a single review and kept where it finds more without reporting safe code.
- An isolated lab that confirms a finding by running a test against your own code, with no network and synthetic secrets. A finding is called confirmed only when that test ran.
- Findings explained in the context of your code; every fix with a regression test, on a branch you merge.
- The basic local review stays free.
- 0.2
The engine and the harness
Planned
- Local Quench runs llama.cpp itself: context, KV cache and GPU layers set from the app, fixed by the Doctor.
- Task replay, and two models side by side on the same task.
- Published comparisons with other agents on the same model and tasks.
- An experiment with an optional adapter that teaches an open 7B–14B model to use tools better.
- 0.3 and on
Security as you work
Planned
- Quench Protect checks the files you change as you work, and optional security gates run in CI.
- A Verified track next to the Community results on the leaderboard.
- 1.0 and on
Teams
When organisations ask for it
- Central policies, an approved-model list, audit logs and a self-hosted control plane.
- Team administration never needs your code or chats on our servers.
What does not change
- Your code and your chats stay on your computer. Anything shared, a benchmark result or a chat, is shared only after you said yes, and you can see what is sent.
- Local Quench for one person stays free, and so do QuenchBench and the basic Quench Protect.
- Security fixes are never held back for a feature.