Fixed candidates
Choose, confirm, or rate.
Ask a choice, yes/no, or score question over a fixed answer set. Candidate keys and score indexes come from your schema, so downstream code does not need an output parser.
Jet
Local structured decisions
Local decisions, structured answers
Give a local language model context and fixed candidates. Jet scores them and returns a result you can use directly.
What Jet does
Score candidates. Keep the schema. Stay local.
Local inference
Run where the data already is.
Use GGUF models on CPU by default, or offload to CUDA, Metal, or Vulkan. With vision enabled, score the same structured questions against PNG or JPEG inputs—including the Doom demo’s screenshot-driven controls.
CLI and Rust API
Stream JSONL or call from Rust.
Send one request per line and receive ordered, machine-readable results from the CLI, or embed the decision engine with typed requests, responses, configuration, and error handling.
Why Jet
Structured answers.
Fixed candidates in. Predictable JSON out.
Fixed candidate scoring
Choose an answer, make a yes/no decision, or rate an item without parsing free-form model output.
Local inference
Run GGUF models on CPU, CUDA, Metal, or Vulkan without standing up an HTTP service.
Streaming JSONL
Process one request per line and receive ordered, machine-readable results from the CLI.
Rust API
Embed the decision engine directly with typed requests, responses, configuration, and error handling.
Multimodal · Doom demo
Score actions from what the agent can see.
The Doom controller in demos/doom shows Jet’s multimodal path: choose among nine controls from a screenshot plus visible status, while the game keeps running. It is a visual control demo, not a trained Doom policy, and visual decision accuracy has not been quantitatively evaluated.
Nine admissible controls
Forward, backward, strafe, turn, fire, use door, and wait—scored as fixed candidates on every observation.
160×100 RGB screenshots
The controller downsamples the live 640×400 view to an exact 160×100 PNG before Jet scores the image.
Visible state only
HUD health and ammo, the previous action, and recent motion cues are included. Hidden enemy positions and map geometry are not.
Vision path, thinking off
The demo uses Jet’s multimodal judge with a vision-language model and projector. The current implementation requires reasoning disabled.
Model comparison
Full MMLU on consumer GPUs.
Accuracy versus complete-process throughput on RTX 4090, from the Jet technical report. Hover or focus a point for exact values. Peak accuracy: Qwen3.5-27B (Vulkan) at 85.01%.
Get Jet
Build it once. Run it locally.
Pick your OS and backend. Commands below match the repository getting-started and advanced guides—clone with submodules first, then install mise 2026.10.0 or newer.
Build and run the CPU CLI
mise install./scripts/download-test-model.shmise exec -- cargo build --release -p jet-cli./target/release/jet judge --model-path models/Qwen3-0.6B-Q8_0.gguf --input tests/fixtures/decisions.jsonl --output -CPU is the default backend. Clone with --recurse-submodules so the pinned llama.cpp submodule is present.