SyzygyResearch/Mach-1-Additive-35B-Multimodal-GGUF

Mach-1-Additive-35B — GGUF

GGUF build of Mach-1-Additive-35B for local inference on consumer hardware (NVIDIA CUDA, Windows/Linux Vulkan, CPU; on Apple Silicon the codec runs on CPU — use Mach Studio for GPU inference on macOS).

This GGUF requires the Mach-1 fork of llama.cpp: SyzygyResearch/llama.cpp-mach1. Mainline llama.cpp will not load it — the model uses custom trellis-coded tensor payloads and decode ops that only exist in the fork.

Files

File	Size	Notes
Mach-1-Additive-35B.mach1.gguf	7.87 GB	text model, payload v3 (additive)

Quick start

Build the fork — pick the backend for your hardware (NVIDIA → CUDA, AMD/Intel → Vulkan):

git clone https://github.com/SyzygyResearch/llama.cpp-mach1
cd llama.cpp-mach1

# NVIDIA (requires the CUDA toolkit)
cmake -B build -DGGML_CUDA=ON
# AMD / Intel / other (requires the Vulkan SDK, incl. glslc)
cmake -B build -DGGML_VULKAN=ON

cmake --build build --config Release -j

Run directly from Hugging Face:

./build/bin/llama-cli -hf SyzygyResearch/Mach-1-Additive-35B-GGUF

or from a local download:

# interactive chat
./build/bin/llama-cli -m Mach-1-Additive-35B.mach1.gguf

# single-turn / scripted use (without -st, llama-cli enters interactive
# conversation mode after answering the prompt)
./build/bin/llama-cli -m Mach-1-Additive-35B.mach1.gguf -st -p "your prompt"

Benchmarks

	Mean retention, 12 benchmarks
Mach-1 Small	96.3%
Ternary Bonsai 27B (PrismML)	93.6%
Gemma 4 Q2_K_XL (Unsloth)	85.6%

Per-benchmark Retention (Score divided by that of the respective full precision model)

Benchmark	Mach-1 Small	Ternary Bonsai 27B	Gemma 4 Q2_K_XL
AIME26	95.8%	92.7%	67.7%
MATH-500	99.2%	98.2%	95.6%
AIME25	93.9%	91.7%	67.2%
GSM8K	99.8%	100.2%	97.3%
MBPP+	98.1%	98.4%	92.2%
HumanEval+	96.5%	98.7%	94.1%
MMLU-Redux	97.0%	94.0%	96.9%
IFEval	94.8%	89.8%	95.5%
MuSR	100.4%	91.6%	91.1%
BFCL-v3	96.3%	98.9%	95.7%
τ²-bench	94.1%	91.2%	73.1%
IFBench	89.8%	77.7%	61.3%
Mean	96.3%	93.6%	85.6%

Mach-1 Small's own scores and teacher scores:

Benchmark	Score	Teacher (Qwen3.6-35B-A3B BF16)	Retention
AIME26	86.25	90.00	95.8%
MATH-500	97.80	98.60	99.2%
AIME25	82.92	88.33	93.9%
GSM8K	95.98	96.21	99.8%
MBPP+	94.18	96.03	98.1%
HumanEval+	91.77	95.12	96.5%
MMLU-Redux	89.88	92.68	97.0%
IFEval	84.42	89.05	94.8%
MuSR	66.93	66.66	100.4%
BFCL-v3	72.19	74.98	96.3%
τ²-bench	74.82	79.51	94.1%
IFBench	58.33	64.97	89.8%

Speed