Edge0-145B
Edge0 open-sourced Edge0-145B and Edge0-120B, sparse models that it says run on a laptop with 29GB peak memory and decode above 80 tokens per second.
Edge0 says the 145B model activates about 10B parameters per step, using a sparse expert design where each group of 64 experts shares two large matrices. Weights come in NVFP4 format on Hugging Face, with an inference framework and deployment tools for macOS, Windows and Linux on Apple M-series and NVIDIA RTX hardware. The company claims coding, reasoning and agent performance that rivals frontier cloud models.
- Date
- Thursday 12 November 2026
- Lab
- Edge0
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| DeepSWE 1.1 | 64.6 Edge0-145B | company |
| Terminal-Bench 2.1 | 86.52 Edge0-145B | company |
| GPQA-Diamond | 90.4 Edge0-145B | company |
| IFBench | 80.6 Edge0-145B | company |
| AA-LCR | 79.3 Edge0-145B | company |
| Mean of six benchmarks | 72.8 Edge0-145B | company |
| Peak memory | 29GB reported | company |
| Decode speed | 80+ tokens/s reported | company |
| Prefill speed | 1,100+ tokens/s reported | company |
All figures are from the company's own blog post and not independently checked. The license was not stated on the page.
Sources
This record was partly confirmed: some claims could not be checked on 13 November 2026. How we check