Mach-1 Additive 35B
Released Aug 3, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- United States
- Modality
- Multimodal
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Benchmarks0 of 8 reported
MMLU-Pro
GPQA Diamond
SWE-bench Verified
Terminal-Bench 2.1
AIME (latest)
Humanity's Last Exam
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- Runs entirely offline on a single 16GB+ Apple Silicon Mac, or streamed straight into a browser tab via WebGPU — no account, API key, or cloud dependency
- Compresses to roughly a tenth the footprint of its teacher through an additive, multiplication-free inference scheme at 1.7 bits per weight
- Reached that compression with under 15 GPU-hours of retraining, far less than comparable extreme-quantization approaches
Weaknesses
- Debut release from a three-person team with no coverage yet on LMArena, Artificial Analysis, or any independent benchmark leaderboard — real-world behavior outside the vendor's own figures is unverified
- Instruction-following degraded more than reasoning or math under compression in the vendor's own retention testing, the softest spot in an otherwise strong profile
- No hosted API at launch — obtaining it means downloading multi-gigabyte weights or running the in-browser demo, not a quick sign-up
Company and model both launched 2026-08-02/03; benchmark figures Syzygy has published are self-reported retention percentages against its teacher model (Qwen3.6-35B-A3B) across 12 evals (e.g. AIME26 99.5%, MATH-500 99.4%, IFBench 83.2% retention), not absolute scores on this site's tracked benchmarks, so all eight benchmark fields are left null. License (Apache-2.0) is inherited unmodified from the Qwen base weights (copyright notice in the repo's LICENSE file still reads Alibaba Cloud). Syzygy's own figures claim up to 120 output tokens/sec on consumer hardware; left out of the speed field since Artificial Analysis has not measured it.
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor