Edge0 open-sources 145B on-device AI model for local laptop inference

Edge0 has open-sourced Edge0-145B and Edge0-120B alongside a unified inference framework and deployment tools. Edge0-145B is described as the first open-source on-device AI model with 145 billion parameters, bringing coding, reasoning, and agent capabilities that rival frontier cloud models. The company says it can run locally on a compatible laptop with 29GB peak memory and decode above 80 tokens per second.

Local performance and design

Edge0-145B is reported to use 29GB peak memory, process prompts at 1,100+ tokens/s, generate at 80+ tokens/s, use approximately 120W typical power, and support 128K to 256K context. Its design shares large matrices across groups of 64 sparse experts, so only about 10B of 145B parameters participate per inference step. Hot experts stay in GPU memory, an expert library of about 40GB sits in system memory, and about 48GB of vocabulary data resides on SSD. The CPU and GPU work in parallel, with speculative decoding verifying multiple candidate tokens at once. Edge0-35B, an earlier phone model, reported peak memory below 2.5GB and 37+ tokens/s on an iPhone 18 Pro Max; those measurements apply to a different model and hardware.

Benchmarks, memory, and local workflows

Edge0-145B scores 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 35.4 on HLE, 90.4 on GPQA-Diamond, 80.6 on IFBench, and 79.3 on AA-LCR. It averages 72.8 across those six benchmarks with 29GB measured peak memory, compared with Qwen3.8-Flash-Next at 71.7 and an estimated 70GB footprint for standard 4-bit deployment. Relative to Qwen3.8-27B, gains include 18.5% on Terminal-Bench and 53.1% on DeepSWE, plus 14.9% on HLE, 1.3% on GPQA, and 1.4% on IFBench. The model can plan, write and run code, debug errors, and inspect results visually. Demonstrations include a Roadster explainer video within an hour, a Manhattan 3D scene in Blender within an hour, a 20-second medieval city fly-through, a drone navigation program, and a PCB design with assembly animation. The company says a Manhattan-scale scene would manually cost about $12,000-18,000 and take at least a month, while a leading cloud model would incur more than $150 in cloud token charges. Local inference incurs no corresponding cloud-model token or API fees, though hardware and electricity still cost money.

Platform, portfolio, and ecosystem

Edge0-145B supports macOS, Windows, and Linux on day one, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090, and RTX 5090. Edge0 released 200 models across four families: Edge0 for complex agent tasks and vibe coding, Audio8 for listening and speaking, Decision0 for fast decisions, and Embedding0 for memory. Its portfolio covers 95% of mainstream chip platforms, spans more than 5 million production devices, passed 15 million downloads in three months, and reached more than 110,000 customer organizations by October 2026. The September Edge0-35B launch post passed 1.2 million views on its first day, and both Edge0 and Audio8 reached No. 1 on Hugging Face's Trending chart. Local inference can keep code and data on the computer, support offline core capabilities, and avoid cloud-model token or API fees as usage grows.

Sources: edge0.ai