- Category
- AI Tools
- Rank
- No. 3522Tools index
Previous survey · No. 3389 ·
- Type
- TOOL
- Use case
- Models: Train & Run
- GitHub
- 15 stars
- Latest release
- v1.0-rc2
- Date
About
PicoLM is a C99 LLM inference engine supporting Llama-2, GPT-2, Qwen 3.6/3.8 (including MoE variants), and Gemma-3n models, built around heavy CPU SIMD acceleration and extreme cross-platform portability, running on everything from DOS and Tru64 Unix to the original iPhone and modern systems. It supports CUDA/HIP GPU acceleration with accelerated IMMA kernels, and version 1.0-rc2 adds a Vulkan backend alongside an expanded set of supported legacy platforms.
At a glance
PicoLM is a small C runtime for serving compatible language models from GGUF files. It is worth evaluating when you want inference on a device with limited resources and can measure how a specific model runs there.
Good fit
Developers testing local inference on small machines or studying a compact runtime.
Consider something else when
Skip it if you need broad model compatibility or a mature accelerator path for your target hardware.
Start here
Check the supported model and quantization list for your target. In a trial, use one small compatible model and measure time to first output and peak memory on the actual device.
Cost and requirements
MIT-licensed C runtime. You need a compatible build target and model weights. The model’s own licence, storage and memory requirements are separate.
Limits to know
A small runtime binary does not describe the memory required for weights and context. The project marks its Vulkan backend preliminary, so use target-specific measurements before planning around it.
Sources checked 2026-11-05. Source-based assessment.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.