Kinetic 0.1.0

The first public release of Kinetic: an open-source engine that runs the same GGUF files as llama.cpp, faster, and none of its speed-ups changes a single token of the answer.

Install

Windows (PowerShell):

irm https://raw.githubusercontent.com/Veloresearch/kinetic/main/scripts/install.ps1 | iex

Linux:

curl -fsSL https://raw.githubusercontent.com/Veloresearch/kinetic/main/scripts/install.sh | sh

Then kinetic picks a model that runs well on your computer and opens a chat. Or download an archive below: cuda runs on NVIDIA through CUDA, other GPUs through Vulkan, and the CPU; cpu is the smallest download. Check it against SHA256SUMS.

Python: pip install kinetic-engine, or pip install llama-cpp-kinetic to run code written for llama-cpp-python unchanged.

What's in 0.1.0

• RAD-1, Warm Start, Zero Wait, Shortlist, Torque, Burst, Convoy and Flywheel, all exact and on by default; --off switches any of them off to measure it
• Receipts: --receipt writes one for an answer, kinetic verify-receipt reruns it and checks every token
• kinetic chat with Markdown, images for Gemma 4 and Qwen3.5, and Zero Wait while you type
• kinetic serve with the OpenAI, llama-server and Ollama APIs on one port, tool calls, JSON Schema and GBNF
• llama.cpp's flags, LoRA adapters, kinetic quantize and kinetic perplexity
• Qwen3.5, Qwen3, Qwen2.5, Llama 3, Gemma 4, Gemma 3, LFM3, LFM2, Granite and more

Measured against llama.cpp b12001 on one laptop (Ryzen 5 6600H, RTX 3060 Laptop 6 GB): rewriting a file runs 4.6x faster on the CPU and 4.5x on the GPU, and the first token after pasting a 22,631-token document arrives in 0.13 s instead of 10.95 s. Short prompts on the GPU are still slower than llama.cpp. Every number and how to reproduce it: https://veloresearch.com/research/kinetic

The binaries are not signed yet: on Windows, SmartScreen may ask you to choose More info, then Run anyway.