Most aggressive quantization schemes trade meaningful capability for efficiency. Mach-1 Additive 35B takes a different point in the design space: by using an additive ternary quantization approach it preserves a large fraction of a BF16 Qwen3.6-35B-A3B teacher's performance while shrinking model footprint and improving runtime efficiency.
Key Capabilities
- High capability retention: reports a mean retention of ~95.0% across 12 diverse benchmarks, often matching or closely approaching the BF16 teacher on math, coding, and reasoning tasks.
- Additive ternary quantization: uses a ternary (three-value) additive quantization scheme to reduce weight precision and memory without wholesale architecture changes, enabling lower memory and potentially faster inference on constrained hardware.
- Practical performance profile: retains near-teacher scores on benchmarks like AIME26, MATH-500, GSM8K, and HumanEval+ while showing clear throughput/latency advantages compared to full-precision variants.
Who it's for and trade-offs
Great fit if you need a large (≈35B) LLM with much lower memory/compute for inference or hosting constraints, but still require strong performance on math, coding, and multi-task benchmarks. It is also useful for research into quantization strategies that prioritize capability retention.
Look elsewhere if you need absolute top-tier BF16 or fp16 accuracy on the hardest long-context reasoning tasks, require certified quantization guarantees for specialized hardware, or depend on architectures explicitly designed for multimodal inputs; some edge-case benchmarks (e.g., certain instruction-following or very long-horizon tasks) show larger retention gaps.
Where it fits
Compared to other compact/quantized 30–35B-class variants, Mach-1 Additive 35B emphasizes capability retention over maximal parameter compression. If your priority is strict minimum memory regardless of task accuracy, ultra-low-bit bespoke solutions may compress further at greater performance cost.
Notes: model page lists Apache-2.0 license and benchmark breakdowns (mean retention 95.0% vs. listed BF16 teacher).
