LLMOS Native is a from-scratch, bare-metal operating system research project whose sole purpose is running local LLM inference and multi-agent execution directly on hardware. No Linux, no Python, no Ollama/vLLM wrapper underneath.

What's here

All of the actual project — source, docs, build system — lives in LLMOS-Native-P1/. See LLMOS-Native-P1/README.md for the full write-up; the short version:

• A C core (model, tensor, context, scheduler, agent executive) targeting x86-64. (An earlier revision also built for AArch64 QEMU and Raspberry Pi 4; that's on hold to concentrate validation effort on one architecture — see docs/ROADMAP.md.)
• An 8 MiB paged tensor arena with immutable shared model weights, copy-on-write KV contexts, capability-scoped agents, and transactional workspace writes with rollback.
• A legacy VirtIO-blk (PCI) driver, a SHA-256-validated LMOF model package format, a byte tokenizer, and Q16.16 fixed-point RMSNorm / RoPE / grouped-query attention / SwiGLU / sampling kernels — a real (not simulated) forward pass, since the kernel build disables x87/SSE entirely and has no runtime float.
• A serial console shell (llmos>) for poking at all of the above: infer, sessions, fork, sched, agent demo, workspace, journal, blk, modelload, infer2.

Read docs/ARCHITECTURE.md and docs/ROADMAP.md inside LLMOS-Native-P1/ before extending the system.

Quick start

cd LLMOS-Native-P1
make test        # build the x86-64 target + host integration tests
make x86 && ./scripts/run-x86.sh

Status

NativeLM-test (the P1 scheduler demo) is a deterministic operator/scheduler test backend, not a trained model. The VirtIO/LMOF/fixed-point-transformer path (blk / modelload / infer2) loads a small synthetic, untrained int8 test package and runs a genuine forward pass — verified working end-to-end in QEMU on x86-64, the only target currently maintained.