Skip to content

Run AI models
on your own computer.

OpenJet sets up and runs open models locally, with a custom engine for supported GPUs (1.5× faster than llama.cpp). Work on code in your terminal, or add local AI to your app with the Python SDK.

Get started See how it works
MTP draftsVerify togetherAccepted tokens
Speculative decoding: draft, verify, acceptMTP drafts several tokens sequentially. The main model verifies the draft positions together and returns the accepted prefix. The green segments illustrate accepted tokens; the lanes are a schematic.
Several draft tokens checked in one GPU pass.

Install OpenJet

pipx install open-jet
openjet setup
openjet
Installation guide

From model setup to useful local work.

Read the docs

Find a model that fits your machine.

Setup checks your GPU and memory, helps you choose a model, downloads it and prepares its runtime.

~/project / openjet setup

$ openjet setup

  1. 1 Detect hardware
  2. 2 Choose model
  3. 3 Prepare runtime

Model catalog

  • Qwen3.8 27B Q4_K_M MTP
  • Qwen3.5 9B
  • Use a local .gguf
Illustrative model selectionLOCAL RUNTIME
Interface concept · choices depend on your hardware.Setup guide

For apps you ship: openjet project provisions the model at build time.

SDK quickstart

Measured on Qwen3.8-27B.

Compare generation speed, prompt processing and output fidelity on the same RTX 3090.

See test conditions

Python-function prompt · greedy generation · 512 tokens · 4 drafts

Across four greedy prompts: 106.7 tok/s mean. Reasoning mode over HTTP: 78.0 tok/s. Speed varies with the workload.

139 tok/s coding result1,595 tok/s prefill

Qwen3.8-27B Q4_K_M · published repository measurements. Generation, prefill and fidelity use different test workloads.

Custom engine requirements

Supported NVIDIA RTX 30, 40 and 50-series GPUs with at least 24 GB VRAM (sm_86, sm_89, sm_120). Linux and CUDA 12+ are required; RTX 50-series needs CUDA 12.8+. Setup builds the engine for the catalog’s Qwen3.8-27B Q4_K_M model. Other models can use llama.cpp.

Full compatibility details

A cloud model plans.
Your local model does the work.

Codex or OpenRouter handles planning and final review. OpenJet delegates repository work, edits and tests to your local model.

Explore Slipstream Connect a provider
openjet · slipstream

$ /mode slipstream

Task: add validation to the import command

CLOUD · Plan and review

Define the expected behaviour and acceptance checks.

  1. Validate required fields
  2. Handle invalid types
  3. Add tests for edge cases

← Concise result for final review

LOCAL · Implement
  • 01 Inspect relevant files
  • 02 Apply changes
  • 03 Run project checks
Example working filessrc/commands/import.py
tests/test_import.py
Cloud: planning & reviewLocal: repository work
Interface concept · example task

Targets a 20% cloud-token share; actual usage varies. Cloud-routed content leaves your machine.

Before you install.

Which GPUs support the custom engine?

Qwen3.8-27B Q4_K_M runs on supported RTX 30, 40 and 50-series GPUs with 24 GB or more VRAM. The custom engine requires Linux and the CUDA toolkit. See exact requirements

Can I use another model or runtime?

Yes. OpenJet can run other GGUF models through llama.cpp, or connect to an OpenAI-compatible server such as Ollama or LM Studio. The custom CUDA engine’s results apply to its supported Qwen3.8-27B configuration.

Does it work offline?

Local inference and the terminal agent work offline once the model and runtime are installed. Initial setup needs a download. Slipstream uses a cloud provider, so that mode requires a connection and sends the routed content to that provider.

How do I embed a model in my app?

Run openjet project at build time to choose and download a model for your target device. Bundle the model with your application and call it through the Python SDK. Models are never downloaded at runtime. SDK quickstart

What does OpenJet cost?

OpenJet is open source under AGPL-3.0. Running a local model does not require an API subscription. You supply the hardware; Slipstream’s cloud model may require a subscription or API credits from your chosen provider.