Edge0 ships a 145B open-weight model that runs in 29GB of memory
Edge0-145B activates roughly 10 billion of its 145 billion parameters per inference step, enough to run on a single workstation GPU or an Apple Silicon Mac.
Edge0 released Edge0-145B on 12 November, an open-weight mixture-of-experts model with 145 billion total parameters, of which only around 10 billion activate per inference step. Peak memory use is 29GB, low enough for a laptop or a single RTX 4090 or 5090, and Edge0 says it processes prompts at over 1,100 tokens a second and generates at 80 or more. Context runs from 128K to 256K tokens, and a smaller companion, Edge0-120B, shipped alongside it.
On Edge0’s own benchmarks it scores 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next’s 58.7, and 80.6 on IFBench, ahead of DeepSeek-V4-Flash-0731’s 79.2. On GPQA-Diamond it scores 90.4, close to DeepSeek-V4-Flash-0731’s 90.8. Weights are on Hugging Face, the inference code is on GitHub, and a desktop app runs the model offline with no account or usage quota.
Why it matters: this is a model that competes with hosted alternatives on hardware you already own, with no token bill and no data leaving the device. For anyone weighing a local model against an API call for coding or agent work, it’s a credible new option to test.
The caveat: every number above is Edge0’s own benchmark, run against its own choice of comparison models, with no independent verification yet.