Mach-1 Additive 35B

Summary

Syzygy product launch. Compresses Qwen 3.6 35B into ~1.7 bpw additive/trellis weights (~7 GB, up to 120 tok/s on consumer laptops). Vendor claim: 95% mean retention vs the BF16 teacher across 12 agentic and reasoning benches (vs Ternary Bonsai 93.6% and Gemma 4 Q2_K_XL 85.6% on the same set; some comparator rows are borrowed). Unlike BitNet, they claim under 15 GPU-hours of retraining and a path to 3T-scale compressed models. Browser demo plus Mach Studio desktop (Apple Silicon beta) and an agent harness. Weights on HF SyzygyResearch/Mach-1-Additive-35B (Apache-2.0).

Keywords

mach-1 · syzygy · quantization · additive · qwen3.6 · local-inference · x

Topics

multiplication-free inference, extreme quantization, local agents

Research notes

• Primary: HF model card (Apache-2.0). Discord/X https://x.com/syzygyeng/status/2084350792841195992 via fxtwitter. Product https://withsyzygy.com. Vendor benches; some comparator rows borrowed. Packed trellis codecs, not GGUF (llama.cpp fork “later this week” at check). Open weights, not a new hosted corpus, so no datasets_local row.