Accessible. Efficient. Honest. ZMLabs builds software that runs serious AI on the hardware people already own — a laptop, an integrated GPU, a browser — instead of a datacenter. The lever is memory and reuse, not a bigger chip.
We publish the numbers behind every product, each one a real measurement with full conditions and its proof file attached → github.com/zmlabs-ai/benchmarks. Capacity is never reported as speed; single runs are labelled n=1; anything we can't stand behind is listed with its reason.
| What it does | A measured figure (full conditions in the proofs) | |
|---|---|---|
| 🧠 memown · memown.com | Run a large local AI on a small machine | A 120B-class MoE boots, serves and decodes on an 8 GB laptop GPU at ~1.5 tok/s (n=1, NVMe-bound — capacity, not speed) → proof |
| 🎮 diciz · diciz.com | Real AI in your browser — no cloud, no install, no account | 31.2 tok/s for Qwen2.5-7B on-device via WebGPU (RTX 4070 Laptop, n=1) → proof |
| 🛟 VRAMPilot · vrampilot.com | Fit any model to your GPU; recover from out-of-memory instead of crashing | A forced 262144 → 131072 context OOM-recovery that keeps serving (RTX 3070 8 GB, e2e, no mocks) → proof |
| 🖼️ GlassBreakr · glassbreakr.com | More FPS through frame generation, on your PC's idle graphics chip | Up to 8× output frames (1 captured + 7 AI-interpolated in 15.8 ms @1080p; ~2× responsive — disclosed frame-gen) → proof |
A benchmark without its conditions is marketing. Every figure we publish carries its model, quantization, hardware, sample count and warm/cold state — and a link to the raw file it came from. "Stand behind" means measured, not believed.
— ZMLabs · Sète, France · contact.zmlabs@proton.me