One engine, every model
Every architecture token on HuggingFace — 569 of them — resolves to one binary. 326,992 checkpoints map to a single 1-bit engine. This is how.
The claim
Go to HuggingFace. Pick any text-generation model. Any architecture — Llama, Qwen, Mistral, Gemma, Zamba, Jamba, Cohere, Falcon, DeepSeek-MoE, MiniMax, BitNet, T5 — and the 1bit.MONSTER engine can run it. Not a wrapper. Not a per-model build. One binary that auto-detects the architecture from the file and dispatches to the right backend.
The daily census keeps the claim honest: 569 architecture tokens, 2,030 HF arch strings, 99.89% HuggingFace coverage, 326,992 checkpoints mapped. A workflow re-checks the numbers every night and updates the site when they move. That number is live, not a marketing figure.
Why it works: the token catalog
Every HF architecture string — llama3, qwen2.5_moe, zamba2, all 2,030 of them — normalizes down to one of 569 architecture tokens. The token captures what actually changes between model families: attention variant, MoE routing, norm placement, RoPE scheme, tie weights. Everything else is shared.
A new model family lands on HF → the daily new-model watcher drafts the mapping → the engine gets the token → the site's numbers update. The gap between a model appearing and the engine running it is measured in days, not quarters.
One binary, every format
The engine reads whatever you give it: GGUF (native), Q4NX 1-bit packed, INT8, TQ2 packed, and raw safetensors. One binary, six backends: NPU, HIP/ROCm, CPU, Vulkan, and more. The gguf_to_q4nx path converts a 4-bit GGUF to a 1-bit pack; the converter is part of the same binary.
| Format | Read | Convert | Typical use |
|---|---|---|---|
| GGUF | native | — | off-the-shelf models |
| Q4NX 1BP | native | from GGUF | our 1-bit pipeline |
| INT8 | native | from GGUF | big-batch CPU |
| TQ2 | native | — | compressed weights |
What 1-bit buys you
A 70B-class model packed to 1-bit fits in the memory of a laptop with room to spare. The engine's measured path on the Ryzen AI NPU went from 244 ms/tok to 3.4 ms/tok across the optimization sprint — that's a 70× improvement on the same silicon, no new hardware.
326,992 arch-bearing checkpoints resolve to 569 tokens, and 569 tokens resolve to one engine. That's the whole pitch.
Try it
build/1bit unified --port 8088 --weights models/ — point it at a directory of GGUFs, hit /v1/models, and pick any of them. The server loads what you ask for, on the backend that's fastest for your hardware, no config file required.