Bonsai-27B-1bit-gguf

Bonsai-27B · 1-bit (Q1_0) GGUF — for the SuperSeed AI Agent

The model the SuperSeed AI Agent app downloads when you switch on Local model (NeuroGrid): PrismML's Bonsai-27B (Qwen3.5 architecture, 27B parameters) in 1-bit Q1_0 — every projection stored as one sign bit per weight with an fp16 scale per 128-weight block. 3.6 GB on disk, 3.6 GB pinned in RAM while it runs.

These are byte-identical redistributions of the files in prism-ml/Bonsai-27B-gguf (Apache-2.0). Credit for the model and the quantization belongs to PrismML.

How it runs

The SuperSeed AI Agent embeds the NeuroGrid engine as a sidecar (neurogrid -solo) and runs this file on the CPU with NEON SDOT kernels written for the Q1_0 block layout (bit=1 → +d, bit=0 → −d), weights mlock'd so they never fall out of RAM, on the performance cores only. On an Apple M5 with 16 GB it loads in ~4 s and decodes at ~3–3.5 tok/s while the laptop stays usable. Everything you type stays on the machine.

Also runs with PrismML's llama.cpp fork (Q1_0 is not in upstream llama.cpp).

License

Apache-2.0, same as the original.