sasori: No-Retrain Multi-Plane Ternarization for GPU-Free LLM Inference

Ternary-weight LLMs (w ∈ {-1,0,+1}, ≈1.58 bits) replace most multiplications with additions and can run with no GPU. But today that path is closed for arbitrary models: it requires either a natively-trained ternary checkpoint (BitNet/TriLM—a handful of models, a "narrow range of applicable model architectures" by the framework authors' own admission), or llama.cpp's single-plane ternary types (TQ1_0/TQ2_0), which recover essentially zero reasoning when used to ternarize an existing full-precision (FP) model post-hoc. We present sasori, a no-retrain pipeline and a new CPU kernel that turn an FP open checkpoint into a high-quality, GPU-free ternary model—quality validated in fake-quant on 14 models across 6+ families (dense and MoE), and end-to-end on CPU on four Qwen-family checkpoints. sasori (i) fits a K-trit-plane decomposition by a joint ridge solve—layer-local, calibration-free, streaming, and embarrassingly parallel across cheap GPUs; and (ii) ships a multiplication-free ggml kernel (new types TQ2P/TQ3P) that runs the result on commodity CPUs. We validate correctness (44 test functions; dequant and reconstruction bit-exact against a reference) and run a real 30B mixture-of-experts (Qwen3-30B-A3B) end-to-end on a pure-CPU host at ≈14 tok/s (single run; the model must fit in RAM to avoid swap). We are deliberately precise about what sasori is not: it is not faster than 4-bit integer quantization on x86 or ARM (decode is memory-bound and a quality-preserving K=3 code is ≈6 bpw, heavier than a 4-bit quant), and it is not a new quantizer (the recovery science is prior work). Its contribution is a categorical capability—GPU-free, retrain-free deployment of arbitrary models—not a throughput or footprint record. We also ship an optional LoRA-KD distillation path that recovers more quality (it brings K=3 to FP on one of two models tested—93% on the other—and rescues the K=2 reasoning collapse) at the cost of a short fine-tune. The pipeline, the kernel patch, and prebuilt binaries are released.