Nanbeige/Nanbeige4.2-3B converted to the CMF format: a single mmap-served file executed by a pure-Rust runtime.

Nanbeige 4.2 is a Looped Transformer: 22 physical layers applied twice (44 virtual layers) with a re-norm at each loop boundary — the engine executes the loop natively — and a reasoning model that thinks before it answers.

Measured on a fanless MacBook Air (Apple M4) — 10-core CPU (4 performance + 6 efficiency), 10-core GPU, 24 GB unified memory, macOS 27. No external cooling, no overclock, nothing plugged in.

Method: cortiq bench --core (greedy, no repetition penalty — the llama-bench contract), median of three runs on an idle machine. The GPU column is native Metal, enabled with CMF_GPU=1. Needs cortiq ≥ 0.5.35 — earlier releases decode this file at ~9 tok/s and take 40 s to ingest a 512-token prompt.

Why a "3B" runs at these rates and not faster. The loop means every token pays for 44 layer passes, ~6.3 G multiply-accumulates — the arithmetic of a 6B dense model. The 166k vocabulary with untied embeddings adds another 1.0B parameters, so the file really holds 4.17B. At q4t that is ~3.9 GB of weights read per token, and an M4 moves about 120 GB/s, which puts a hard ceiling near 31 tok/s on this machine no matter how good the kernels are. 22.4 tok/s is ~72% of that ceiling. A laptop with more memory bandwidth (M4 Pro/Max, or a discrete GPU through the Vulkan backend) scales roughly with it.

CMF_GPU=1 selects native Metal on Apple Silicon; on Linux/Windows the same flag picks the Vulkan/DX12 backend. Without it the runtime stays on the CPU, which is the second column above.

Reasoning models deliberate first — give them a few hundred --max-tokens to reach the final answer.

The converter streams the checkpoint shard by shard, so the free-disk requirement is one shard plus the output.

The looped forward is checked against a float32 PyTorch reference built from the upstream modeling code: same argmax and same top-5 ordering on the same prompt, with the q4t logits within ~2 of the reference's. The model's own quirks survive the quantization intact — including a reasoning opener it is very sure about.

Apache-2.0, inherited from the base model. Weights © Nanbeige LLM Lab; this repackaging only changes the storage format.

CMF Mobile is a Flutter app built on this runtime: install it, point it at this 2.4 GB .cmf and chat on-device with streaming tokens — or flip on its server mode and the phone becomes an OpenAI-compatible endpoint for your network. The engine, converter and format spec live at https://github.com/infosave2007/cmf.