Edge0-145B: a 145B model in 29GB, with a benchmark table from last year
A 145 billion parameter model that peaks at 29GB of memory and decodes at 80+ tokens a second on a laptop. That is the pitch from Edge0, which open-sourced Edge0-145B and Edge0-120B together with its inference framework and an app. The blog post went up Nov 12 at 18:00 UTC, and the launch hit the Hacker News front page that night at 334 points: https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer . Weights are at https://huggingface.co/Edge0 and the code (Apache-2.0, already 10,891 stars) at https://github.com/Edge0-AI/Edge0
The trick is memory tiering, not shrinking. Groups of 64 experts share two large matrices, and the model was trained in its on-device shape from day one. Hot experts sit in GPU memory, a roughly 40GB expert library lives in system RAM, and about 48GB of vocabulary data stays on the SSD and is read on demand. CPU and GPU split the work, with speculative decoding on top. Edge0 claims 1,100+ tokens a second of prefill, around 120W of power, and 128K to 256K of context. The 145B NVFP4 build already has 14,134 downloads.
Now the fine print. Every number in the table is Edge0's own eval. DeepSWE 1.1 at 64.6, Terminal-Bench 2.1 at 86.52, GPQA-Diamond 90.4. The Terminal-Bench comparison is against Claude Opus 4.6 Max at 78.2, a model two generations old, and the other rivals are flash-tier cloud models like Qwen3.8-Flash-Next and DeepSeek-V4-Flash. The technical report is still being released. So the honest reading is that Edge0 roughly matches last year's cloud coding models on a laptop, by its own measurement.
That is still a big deal for agents. A coding agent that runs locally pays no per-token bill, sends no code to anyone, and keeps working on a plane. The cost-per-task argument that has dominated this month mostly happens inside cloud pricing. Edge0 is betting that for a lot of agent loops, the cheapest token is the one that never leaves your machine. Wait for independent runs before trusting the leaderboard, but the memory architecture is the part worth stealing either way.
← Back to all articles
The trick is memory tiering, not shrinking. Groups of 64 experts share two large matrices, and the model was trained in its on-device shape from day one. Hot experts sit in GPU memory, a roughly 40GB expert library lives in system RAM, and about 48GB of vocabulary data stays on the SSD and is read on demand. CPU and GPU split the work, with speculative decoding on top. Edge0 claims 1,100+ tokens a second of prefill, around 120W of power, and 128K to 256K of context. The 145B NVFP4 build already has 14,134 downloads.
Now the fine print. Every number in the table is Edge0's own eval. DeepSWE 1.1 at 64.6, Terminal-Bench 2.1 at 86.52, GPQA-Diamond 90.4. The Terminal-Bench comparison is against Claude Opus 4.6 Max at 78.2, a model two generations old, and the other rivals are flash-tier cloud models like Qwen3.8-Flash-Next and DeepSeek-V4-Flash. The technical report is still being released. So the honest reading is that Edge0 roughly matches last year's cloud coding models on a laptop, by its own measurement.
That is still a big deal for agents. A coding agent that runs locally pays no per-token bill, sends no code to anyone, and keeps working on a plane. The cost-per-task argument that has dominated this month mostly happens inside cloud pricing. Edge0 is betting that for a lot of agent loops, the cheapest token is the one that never leaves your machine. Wait for independent runs before trusting the leaderboard, but the memory architecture is the part worth stealing either way.
Comments