Your GPU is more capable than you think
Demand for sovereign intelligence will rapidly outpace bare-metal access to compute. Syzygy shapes the intelligence you need, at the speed you need, for the hardware you rent or own.
Introducing
Mach-2 Additive Medium
125.7 billion text parameters on a single 32 GB GPU. Qwen3.8-Flash-Next at 1.70 bits per weight, inferencing solely with addition rather than multiplication, in 26.7 GB while retaining 96% of the BF16 model's benchmark scores.
- 1.70 bits
- per weight
- 125.7B
- text parameters
- 96%
- average retention vs BF16
- 26.7GB
- on disk
Our mission
We expect hardware to always be one step behind software. Each generation of models asks for more memory and compute than the available chips can supply, so compute remains the leading constraint on what intelligence can do, and on who gets to use it.
Inference began with closed weights on closed hardware: you rented both the model and the machine. Open-weight models changed the first half, but most of them still run on someone else's GPUs. The next step is open weights on open hardware, where you control the hardware and the deployment.
Owning the hardware only pays off if you can get the most out of it. We compress and reshape frontier models so more intelligence fits in the memory you have, and we build the runtime alongside them so the same GPU serves more work. We are extending that to the GPU itself: reusing and forking its state, and freezing, thawing, and moving workloads between nodes, so limited capacity goes further.
Our mission is to maximize intelligence per dollar on any hardware you own or rent.
Syzygy
/ˈsizəjē/ · noun
Co-designing the model with the memory it runs in
Two limits decide what a GPU can do with a model: how much fits in memory, and how fast the weights move through it. We design the model and the runtime together against both.
Memory capacity
Fit more intelligence in the memory you own
Compress the model's weights to fit the available memory while preserving task quality.
On a MacBook M5 Max: Mach-2 Additive Medium takes Qwen3.8-Flash-Next from 251.4 GB to 26.7 GB and keeps 96% of its benchmark score.
Memory bandwidth
Move fewer bytes for each useful token
Choose model formats and runtime kernels together so the GPU spends less time moving weights.
On NVIDIA B300s: Mach DeepSeek V4.1 Flash serves each user 2.2x faster than stock vLLM at 51% of the cost, with 16 coding agents running on one 8-GPU node.
Join us
We are a team of mathematics and physics researchers working on model compression and inference runtimes. If our mission resonates, please reach out.