Get started

Hardware and models

litecode runs on any machine, from a 4 GB laptop to a workstation with a large GPU. It measures the machine and picks models that fit. On a small PC it picks small, fast models. On a big machine it picks bigger, stronger models.

How litecode chooses#

litecode setup reads total RAM, CPU cores, disk type, and GPU.
It plans with about 60% of total RAM, so your browser and editor keep working.
It picks a model for each role: code, chat, fim (autocomplete), security, and quality (hard tasks).
With a GPU build, it moves model layers to the GPU (gpu_layers = 99).

Tiers#

Machine	What litecode loads	Expected use
4 GB RAM, no GPU	Qwen2.5-Coder-1.5B + 0.5B autocomplete	code help, explanations, small fixes
8 GB RAM, no GPU	MiniCPM5-2B + 1.5B autocomplete	daily coding, security review, writing
16 GB RAM, no GPU	LFM2.5-8B-A1B (fast MoE) + security model on demand	full agent work on real projects
32 GB RAM	adds Qwen3.5-9B as the quality model	harder multi-file tasks
GPU, 12 GB+ VRAM	Qwen3.6-35B-A3B and other 30B-class models	large refactors, deep reviews, long context

These are defaults. Any GGUF model works. Set it with litecode config roles.code MyModel.gguf.

GPUs#

GPU	Engine build	How to get it
NVIDIA	CUDA	setup offers it when it finds `nvidia-smi`
AMD, Intel, NVIDIA	Vulkan	setup offers it when it finds `vulkaninfo`
Apple Silicon	Metal (in the macOS build)	automatic
Old GPUs (2 GB, Kepler era)	CPU	litecode ignores the GPU

Set how many layers go to the GPU: litecode config gpu_layers 99 (all) or a smaller number to share VRAM.

What makes it fast#

Generation speed depends mostly on memory bandwidth. Two RAM sticks (dual channel) give about 2x the speed of one.
Models on an SSD load in about 5 seconds. On a hard drive they take about 30 seconds. litecode keeps the model loaded between tasks.
The prompt cache is saved per project. The next session skips reprocessing the long system prompt.
N-gram speculative decoding speeds up edits that repeat existing code.
litecode tune measures all of this on your machine and keeps the fastest settings.

Reference machine#

All published litecode numbers come from this low-end PC, so you can compare:

Part	Value
CPU	AMD Ryzen 3 3200G, 4 cores, 4 threads (2019)
RAM	16 GB DDR4-2133, dual channel
GPU	GeForce GT 710 2 GB (too old, not used)
Disk	7200 rpm hard drive for models

If litecode runs well on this machine, it runs better on yours. See Benchmarks.

Source: docs/HARDWARE.md