Introducing Mach 1 Additive
A 35 billion parameter model that runs inference without ever multiplying by a weight. At 1.7 bits per weight, it retains 96.3% of the performance of its full-precision teacher, Qwen 3.6 35B, across 12 agentic and reasoning benchmarks — at a tenth of the size.
At ~8 GB, Mach-1 fits comfortably on consumer laptops, making local inference useful, not just feasible.
Capability, not just compression
Score kept as a percentage of each model's own BF16 teacher, measured on the same harness across knowledge, math, coding, instruction-following, and agentic benchmarks.
Dot plot of student-to-teacher benchmark score ratios across all 12 benchmarks, for all three models. Mach 1 Additive 35B averages 96.3%, Ternary Bonsai 27B 93.6%, and Gemma 4 Q2_K_XL 85.6%.
Mach 1 Additive 35B
96.3%
Mean over 12 benchmarks
Bonsai 27B
93.6%
Ternary
Gemma 4 Q2_K_XL
85.6%
2-bit GGUF
Bar chart of local wall-clock time to answer. For a 128-token prompt Mach 1 Additive 35B takes 3.7 seconds, Bonsai 27B 6.1, and Gemma 4 Q2 11.5. For a 2,048-token prompt Mach 1 Additive 35B takes 5.5 seconds, Bonsai 27B 9.8, and Gemma 4 Q2 16. For a 8,192-token prompt Mach 1 Additive 35B takes 12.9 seconds, Bonsai 27B 22.3, and Gemma 4 Q2 36.8.
Prefill plus decode wall-clock time in seconds; lower is better.Apple Silicon local run, July 2026
Two bar charts of local intelligence speed. Mach 1 Additive 35B scores 15.6 pts/s and 1.97 pts/GB/s; Bonsai 27B scores 9.2 and 1.28; Gemma 4 Q2 scores 4.9 and 0.42.
Intelligence per second
pts/s
Intelligence density per second
pts/GB/s
Benchmark-score composite per second, and per GB of model memoryApple Silicon local run, July 2026