Edge0's 145B 'open' model 'runs on a laptop' — independent check finds the memory and open-source claims overstated

AI team Edge0 (company AutoArk) released two large language models, Edge0-145B and Edge0-120B, plus its own inference framework on November 12, calling the 145B model "the first open model of this size that runs on a personal computer": 145 billion total parameters with about 10 billion active per token, 29GB of peak memory, decoding at more than 80 tokens per second and a 128K–256K context window, aimed at laptops such as a 48GB MacBook Pro. The weights are on Hugging Face and the code on GitHub under Apache 2.0.

29GB is only the GPU peak; the card measures about 61.5 GiB in total

Independent verification site Ground Truth reported on November 13 that the blog headline's 29GB is only the peak on the GPU side. The model card's own measured figures for the NVFP4 build are 23.25 GiB of VRAM plus 38.27 GiB of host RAM — about 61.5 GiB in total — and the weights themselves are 47.79 GiB for NVFP4 and 40.81 GiB for the Q3_K build. The blog itself says about 40GB of expert weights sit in system RAM and about 48GB of vocabulary data on SSD. The speed figures are desktop-GPU numbers too: 79.6 tokens per second was measured on an RTX 5090 D, and the Apple M2 Ultra row at 60.9 tokens per second carries a note that macOS release availability is not established, while the engine documents CUDA, not Mac.

The base is Alibaba's Qwen, and the licence is not standard open source

The blog says the model is trained "in its on-device architecture from the outset", but the Hugging Face metadata names Alibaba's Qwen3.8-Flash-Next — a roughly 180-billion-parameter model released in August — as the base model. Edge0's own release-scope file acknowledges the actual training base revision is unverified and claims no ownership of Qwen materials. On licensing, the weights use a custom Edge0 Community Model License 1.0: selling the weights, running a paid model service for third parties, or offering a standalone AI coding or office product all require a separate written licence — not a standard open-source licence.

All benchmarks are self-tested; the technical report is not out

Every benchmark score comes from Edge0's own evaluations: 64.6 on a coding-agent test against 58.7 for its Qwen base, 86.52 on a terminal test, and a six-test mean of 72.8 against 71.7. But Edge0 trails Qwen on GPQA-Diamond (90.4 against 91.7), on an instruction-following test and on Humanity's Last Exam. The technical report and reproduction instructions have not been published, and Hacker News commenters called the headline "a flat lie", arguing a 48GB laptop could not cope with a roughly 70GB combined footprint. Until someone independently measures a laptop run, the story is the gap between a 29GB headline and a measured 61.5 GiB total.

(Illustration: a large language model on a laptop.)