FAQ

What is 1bit systems?

The 1-bit inference engine. A Rust orchestration stack running Microsoft's BitNet-b1.58-2B-4T on an AMD Strix Halo mini-PC with native HIP kernels. OpenAI-compatible API, MCP server, 17 specialist agents, live landing page. Bring any OpenAI-compat client, plug it in, the agents keep the stack running while you work.

What does it run on?

AMD Ryzen AI MAX+ 395 (Strix Halo) mini-PC, $2-3k. 128 GB LPDDR5 unified memory, Radeon 8060S iGPU (gfx1151, RDNA 3.5), 256 GB/s bandwidth, <150 W under load. CachyOS for the kernel-7 NPU driver support.

How fast?

• 83 tok/s at 64-token generation
• 68 tok/s at 1024-token generation
• Perplexity 9.18 on wikitext-103 (matches Microsoft paper's 9.1607 within ±0.05)
• 96.66% byte-identical output vs the reference C++ server after ~10 000 side-by-side rounds

How do I install it?

git clone git@github.com:bong-water-water-bong/1bit systems-core.git
cd 1bit systems-core
./install-strixhalo.sh

One script, ~5 minutes on a fresh CachyOS box. After: halo doctor, halo chat, halo say "hello".

What clients work with it?

Anything that accepts an OpenAI-compatible base_url:

• Open WebUI — polished desktop-style chat, full RAG + MCP
• LibreChat — yaml-configured OpenAI alternative
• DSPy (Stanford) — compile declarative LM programs against 1bit-server
• Claude Code — 1bit-mcp registers as MCP server; 17 specialists appear as tools
• 1bit-helm — our own native-Rust egui desktop client (renamed from halo-gaia 2026-04-20)
• lemonade-python-sdk via the Lemonade-compat shim on :8200

Why 1.58-bit and not 1-bit?

See Why-Ternary.md. Summary: ternary has a load-bearing zero that maps to activation sparsity, hardware efficiency, and 10× memory reduction with near-zero accuracy loss.

Why Rust and not Python?

See Why-No-Python.md. Summary: 200 ms cold-start vs 10 seconds, 15 MB binary vs 1.2 GB venv, Result<T, E> vs KeyError at line unknown.

Is it ready?

For private-beta testers, yes. PPL parity proven, shadow-burnin at 96.66% byte-exact, CI green, tester-installable. Public launch (reddit post, Steam-community-style rollout) after the 72-hour burnin gate clears. See ../../CUTOVER.md.

Can I run it on Apple Silicon?

Yes. The workspace is feature-gated — cargo build --workspace --features mlx-apple pulls in bitnet-mlx-rs and targets M-series via MLX. Same code, different kernel backend. AMD is the performance target; Apple is supported but not optimized.

Can I run it on NVIDIA?

Not today. Our kernels are HIP. A future CUDA port is technically doable (HIP → CUDA is mostly mechanical) but not planned. A Windows/NVIDIA build would need a new maintainer.

Can I use my own model?

Partially:

• Any BitNet GGUF with IQ2_S weights should load through our 1bit-core::gguf path (commit 2d1ec89). Currently parse-only; the IQ2_S → .h1b bit-unpack lands next sprint.
• Microsoft's bitnet-b1.58-2B-4T works today via our .h1b format + requantizer.
• Non-BitNet architectures (Llama, Mistral, Qwen) don't run on the ternary kernels. Those models use their own non-1.58-bit formats.

What's 1bit-agents?

17 specialist agents running in the background — anvil rebuilds kernels on commit, librarian keeps the changelog, quartermaster triages issues, magistrate scans PRs for secrets + commit-msg compliance. See the 1bit-agents crate. Exposed via MCP so Claude Code and DSPy can call them as tools.

What's shadow-burnin?

See Why-Shadow-Burnin.md. Continuous parity check between the gen-1 C++ server and gen-2 Rust server so we can cut over with evidence, not faith.

Why not just use Ollama / LM Studio / vLLM?

• Ollama — CPU-first, runs BitNet via llama.cpp's ternary kernels, no native gfx1151 optimization. We're ~5× faster on the same box.
• LM Studio — GUI-only, closed-source. Can't drive it from automation.
• vLLM — NVIDIA-first, CUDA-only. No AMD iGPU path today.

Also: none of them ship an MCP server, self-maintaining agents, Lemonade-compat gateway, or a recording-ready landing page. 1bit systems is the batteries-included bundle, not just the kernel.

Is it open source?

• Kernels (bong-water-water-bong/rocm-cpp) — MIT, public.
• Rust monorepo (1bit-systems) — private until launch. Collaborator invites available.
• Everything we fork or borrow is credited in ../../CONTRIBUTING.md.

Who's behind it?

One operator, one box, one Claude Code session that won't quit. See ../../CONTRIBUTING.md for acknowledgements — huge thanks to Light Heart Labs, Microsoft's BitNet team, the llama.cpp/ggml maintainers, and everyone whose shoulders we're standing on.