Picolm

Note: MoE model GPU tensor upload is not yet supported on HIP/ROCm (CUDA works fine). MoE models on HIP fall back to CPU SSM matmul; full GPU upload path planned for next release.

MSYS2 CUDA build target: hunger-msys-direct Makefile target for MSYS2 SSH builds with nvcc+cl.exe

The original PicoLM was a clever proof-of-concept: mmap a GGUF, stream layers through 45MB of RAM, call it a day. Noble effort, but it was missing basically everything you need to actually use an LLM. So I went in and added hundreds of commits across quantization acceleration, model support, HTTP server, GPU backends, and cross-platform fixes.

Since then, progress has been steady. The main goals are (beyond satisfying curiosity on what Qwen-3.6-27B can and can't do):

First, portability. The only good software is one that runs (and is tested) on everything from the last 50 years. PicoLM runs from 32-bit MS-DOS through Raspberry Pi, MIPS/OpenWRT, AMD ROCm (MI50 tested), Metal (if someone bothers), to RTX 4090 or DGX Spark.