KLIF
GitHub
Download
Language models, images, speech, music and video: KLIF starts the servers you already use, shows what each one does to your GPU right now, and stops them again. Nothing hidden, nothing bundled.
Backendsany build
> klif-cli status SYSTEM STATUS MODEL SPEED s1 online gemma 4 26b s2 busy qwen 3.8 27b 32.5 t/s cgi busy qwen image 2.1 12.4 s s3 offline flash-next > klif-cli launch s3 --yes --wait
For people and for agents
JSON with a stable schema, an Agent Skill that teaches the contract in one file, and --yes wherever something starts or stops.
Kinds
6
llm · image · tts · stt · music · video
Platforms
3
Windows · macOS · Linux
Skins
9
full window and compact
Slots
3
conversations per System
License
MIT
free, open source
Weights shipped
0
you bring the models
New in 0.4 · leases
One card, several models. An agent whose own model runs in KLIF hands the GPU to another System for one job and picks its conversation up where it left it, all in one tool call.
> klif-cli lease cgi --yes --from s1 -- python make_sprites.pyexit 0 · everything back
klif-cli lease returns only when everything is back, also when the job fails or the System faults, and its exit code says which. If klif-cli itself is killed, KLIF restores without it.child_failedrestore_failedWhat it does
6kinds
System 1, System 2, System CGI, System TTS, or whatever you call yours. Each one has its own server, session, logs and telemetry, and shows online, busy, starting, offline or fault at a glance. KLIF checks ports, exclusive GPUs and free VRAM before a launch, and never kills a process it does not own.
200Ktokens kept
A llama.cpp or Strata System saves what its server holds, the KV cache of a long conversation, into one of three slots on disk and loads it back, so a 200K-token session is not read again. Every skin has the three slot keys: click loads, holding saves.
1command line
A preset is the launch command: program, arguments, working folder, environment. The Options drawer, klif-cli plan and klif.toml show the same thing. Params such as reasoning on or off, or the image size, replace a folder of near-identical scripts.
2machines, one window
Another computer running KLIF can offer its Systems to this one: opt-in, token-authenticated, off by default. KLIF also keeps your records, the best decode, prefill and time per image each model file reached on each backend, from everyday use.
Nine skins · same data
Every skin draws the same live telemetry, in a full window and in a 960×640 layout. F2 cycles them, and a System can have its own.
The transformer in 3D on an amber phosphor screen: layers, experts, KV cache.
Widescreen · 1280×800 · choose a skin to see its other two layouts above
All nine at once · top: the window · bottom: the 3.5-inch panel
From a launcher script to a stack
KLIF was built for daily work with local models on one machine, and that is still what it is for. Features land when they solve a real problem there.
indevA PowerShell menu: model, quant, backend, GPU, context, port. Press Enter.
one server at a time
0.1The same idea in a window: a tab per model tier, the command and its console.
the name, the mark
0.2A native core and klif-cli, live GPU memory, a panel mode, the first skins.
Tauri, Rust
0.3 · 0.4Any number of Systems, other machines, editable commands, downloads, a CLI for agents, and leases.
Now 0.4.0Platforms
A preset is a plain command, so KLIF launches whatever build of a server you use: CUDA, HIP, Vulkan, SYCL, Metal or CPU.
You bring
The servers you want to run and their model files, in folders you choose. KLIF is not a model zoo and not an inference runtime: it starts them, watches them and stops them. No weights are shipped.