01 · overview
The model runs on your hardware. Full stop.
Peridot is a local AI kernel built around the same principle as the rest of the uncoalesced stack: the model runs on your hardware, under your control, with nothing phoning home. Configuration and system behavior are governed by an explicit, versioned “constitution” file rather than opaque defaults, and the project maintains a running changelog across releases.
The current stable release is v1.6.0. It runs on a new native CUDA engine: the default model is now Qwen3.8-27B, held entirely on an 8 GB laptop GPU, decoding at 23.6 tokens a second where the old 14B default managed 5.3. Answers stream token by token. On top of that engine sits Peridot’s first agentic layer. None of it moves anything off your machine: extensions live in a local folder, and web access stays off until you switch it on.
It runs on Windows and natively on Linux (Debian 12, Ubuntu 22.04+ and Arch).
02 · new in 1.6.0
Agentic, and still sovereign
Skills and plugins
Skills are Markdown prompts, plugins are Python tools, and both live in a local folder. A plugin stays off until you approve it, runs in a sandboxed child process, and loses that approval the moment its files change.
The model uses tools
Approved tools are offered to the model, which can call them itself for up to 8 steps a turn. Each call shows up live in the chat as it runs.
Web search, off by default
A bundled plugin searches DuckDuckGo or your own SearXNG. It is in beta, and Peridot makes no web requests until you switch WEB on.
Sovereign Invocation
Claude Code, Codex, Gemini CLI or any MCP client can hand a job to Peridot. The local model works only on folders you allowlisted, and by default you approve a redacted summary before anything goes back.
A relay that gives the GPU back
A small tray process starts the engine when a delegated job arrives and stops it after 5 idle minutes, so the VRAM returns to Folding@home or anything else.
No refusals of its own
Peridot dropped its refusal list and keyword filter. What a model will or will not say now comes down to the model you load, since alignment trained into its weights still applies.
Two things to know. The plugin sandbox is not enforced by the operating system, so a plugin written to escape it could, which is why you should only approve plugins you trust. And computer use is a scaffold for now: it is switched off, the model is never offered it, and the working version is planned for v1.6.1.
03 · how it works
Air-gapped by architecture, not by policy
A sovereign, local kernel
Everything runs inside the boundary of your machine. Local hardware feeds a sovereign kernel, which drives the user interface and the audio module directly. Inference runs on your machine, web access is off until you turn it on, and delegated jobs read your files locally and return only a result you can review first.

Idle compute, instant handoff
Idle GPU memory is treated as a design defect. While you're away, the kernel puts the hardware to work on background compute like protein folding, and when you return it signals the background work to pause in about 21 milliseconds and has the GPU fully back for inference in under 510 milliseconds.
Deciding when to hand it back is its own problem, and it has its own project inside Peridot: ZAT-SCS reads keystroke acceleration and microphone amplitude at 10Hz to predict that a prompt is coming, and reclaims the card before you press Enter rather than after. Since v1.6.0 it is opt-in and off by default, so no microphone stream or keyboard hook starts at boot.

A versioned constitution
No opaque defaults. The kernel's configuration and system behavior are governed by an explicit “constitution” file that is versioned, diffable and yours to edit. In v1.6.0 it was rewritten in plain language: it tells the model its name, today’s date and exactly which tools it has, and it tells the model never to claim a search it did not run. Every release ships against a running changelog, so behavior changes are visible history.
Measured, on one laptop GPU
The decode figure comes from an isolated benchmark with no HTTP and no cache. The project retracted an earlier, higher figure once it turned out to be measuring cache hits.
Tested against RAG workloads
Retrieval-augmented generation is part of the kernel's testing scope. The “RAG” tag on this project names something it is tested against.
- Decode tokens/s, higher is better
- v1.5: 5.3v1.6.0: 23.6
- Prefill tokens/s, higher is better
- v1.5: ~34v1.6.0: ~500
- First /ask turn s, lower is better
- v1.5: 18.8v1.6.0: 2.2
- Vector search, 100k ms, lower is better
- v1.5: 258v1.6.0: 8.4
04 · stack
Tech stack
| Language | Python 3.11 |
|---|---|
| Inference | llama-cpp-python, source-built with native CUDA kernels, running GGUF models |
| Default model | Qwen3.8-27B (UD-IQ1_S), every layer on an 8 GB GPU, 8k context, streamed |
| Memory | Split tensor allocation across GPU VRAM and system RAM for models that do not fit |
| Extensions | Markdown skills and Python plugins in a local extensions/ folder |
| Agent bridge | An MCP server written on the Python standard library alone |
| Retrieval | TurboVec, a Rust-backed vector index with 4-bit TurboQuant compression |
| History | A local SQLite conversation ledger, no accounts or cloud sync |
| Platforms | Windows, and native Linux on Debian 12, Ubuntu 22.04+ and Arch |
| License | MIT |
The benchmark figures above were measured on an NVIDIA RTX 5050 Laptop GPU with 8 GB.
05 · on github
Development in the open
06 · from the lab
Illustrations
More Peridot illustrations are being produced and will drop into this gallery as they land.
Runs local. Stays local.
Version 1.6.0 is out. Built, shipped, and getting regular updates.

