On-device model company Edge0 open-sourced Edge0-145B and Edge0-120B on November 12, together with a unified inference framework and deployment tools. The company says 145B is the first open-source on-device model to reach 145 billion parameters, and that it targets personal laptops rather than servers.
The launch supports macOS, Windows and Linux, covering Apple M-series chips and Nvidia platforms including RTX Spark, RTX 4090 and RTX 5090. The weights are hosted on Hugging Face and the inference framework on GitHub. There is also a desktop app that needs no account and has no usage cap, and its core features work offline.
Benchmarks target “Flash”-class cloud models
In the results Edge0 published, it scores 64.6 on DeepSWE 1.1, against 58.7 for Qwen3.8-Flash-Next and 54.4 for DeepSeek-V4-Flash-0731. On Terminal-Bench 2.1 it scores 86.52, ahead of the 83.3 and 82.7 posted by those two. Its GPQA-Diamond score of 90.4 is slightly below DeepSeek-V4-Flash-0731 at 90.8, and its HLE score of 35.4 is slightly below Qwen3.8-Flash-Next at 35.9.
These numbers come with a caveat: the results for Edge0, Qwen, DeepSeek and Claude all come from Edge0’s internal evaluation, and the full test environment and reproduction notes are to be released with a technical report. The Claude entry in the comparison table is Opus 4.6 (Max), an earlier generation, and current flagships have long since moved on. Until third parties rerun the tests, the table should be read as vendor-reported.
Compared with the company’s own earlier phone model, the jump is large. Edge0-35B, released in September, uses about 2.5GB of memory on an iPhone 18 Pro Max and runs at more than 37 tokens per second. The 145B model has roughly four times the parameters and is meant for laptops. Edge0 cites Apple’s on-device foundation model at about 3B, and Google’s Gemini Nano in 1.8B and 3.25B versions, to argue that on-device models have generally stayed under 10B.
How 145 billion parameters fit into 29GB
A rough calculation shows the problem: at conventional 4-bit quantization, 145B parameters need about 72GB for the weights alone, far above 29GB. Edge0 describes a four-layer approach.
Shared large matrices in the architecture. Every 64 experts share two sets of large matrices, and each expert keeps only a small set of parameters describing how it differs. This cuts the amount of weight data stored and moved, while the number of experts stays the same. Only about 10B parameters are activated per inference step.
Training built for the device from the start. The model was trained on an on-device architecture from the beginning, rather than following the “train a large cloud model first, then compress and prune it” route.
A three-tier storage pipeline for loading. Frequently used “hot experts” stay in video memory, an expert pool of about 40GB sits in system memory and is moved into video memory on demand, and about 48GB of vocabulary data lives on the SSD and is read as needed. Reading and computing overlap, which avoids serial waiting.
CPU and GPU working together. The GPU computes the experts held in cache while the CPU handles the rest in parallel. Speculative decoding then verifies several candidate tokens at once, spreading the cost of reading weights across more output.
One point the blog leaves unclear: it does not explain how the 29GB peak memory squares with the roughly 40GB expert pool kept in system memory, or whether the figure counts only the video-memory side. Edge0’s own comparison chart labels the model as usable on a 48GB laptop, so users who want to run it locally would be safer planning for 48GB of unified memory.
Demos and company claims
Edge0 showed several local agent demos: producing an explainer video within an hour, building a Manhattan city scene in Blender, writing a drone program that flies through obstacles to find a red backpack, and going from KiCad board layout all the way to a UE5 assembly animation. All of these are internal company tests, and the estimates of how much labor and cloud token cost they save are the company’s own calculations.
The company also says its four model families have released 200 models in a year, that its open-source models were downloaded more than 15 million times in three months, and that it had more than 110,000 customer organizations as of October. These scale figures currently rest on the company’s word alone and cannot be verified.
Sources: Edge0 official launch blog, Hugging Face model page, CocoLoop, Edge0 GitHub inference framework documentation; all benchmark scores are from Edge0’s internal evaluation, and the 29GB figure is the vendor-measured peak memory.