Edge0-145B's Model Card Names Qwen as Its Base, and the Launch Post Never Does

Edge0 released a 145-billion-parameter model on November 12 and called it "the world's first open-source on-device AI model with 145 billion parameters." The launch post never says what the model was built from. Its Hugging Face card does: the base is Qwen/Qwen3.8-Flash-Next, an Alibaba release.

The post, titled "Frontier AI Capabilities on Your Own Computer", describes a sparse mixture-of-experts design with about 10 billion parameters active per inference step. Edge0 said peak memory is 29GB and that the model runs on a laptop such as a MacBook Pro. Qwen appears in the post only as a column in the benchmark tables.

What the model card says that the post does not

The model card for Edge0-145B-NVFP4 lists the base model as Qwen/Qwen3.8-Flash-Next and calls the repository a quantized version of it. It lists 48 layers and 512 experts per layer, with 10 routed experts and 1 shared expert active per token. The context window is 256K tokens, and the card gives active parameters as 10.12 billion.

The licence is two documents. The card names the Edge0 Community Model License 1.0 for Edge0's own contributions and says the Qwen materials stay under the Qwen Community License 1.0, so both sets of terms apply. The launch post calls the release open-source and names no licence at all.

The card does not describe training data. It says only that the model was "trained natively at 4 bits".

How it compares with the Qwen model underneath

Qwen's own Qwen3.8-Flash-Next card gives 125 billion total parameters and 6 billion activated per token, with the same 512 experts and the same 10 routed plus 1 shared split. It also lists 51 billion n-gram embedding parameters and 4 billion for multi-token prediction, which brings the size Hugging Face displays to 180 billion.

Edge0's card describes a 51.2-billion-parameter n-gram embedding table and a 1.26-billion-parameter draft head. Neither card says how Edge0 reached 145 billion from 125 billion, and the Edge0 card's own metadata panel shows "5B params", which conflicts with its text.

1. Edge0-145B145 B
2. Qwen3.8-Flash-Next125 B

Source: Hugging Face model cards for Edge0-145B-NVFP4 and Qwen3.8-Flash-Next, accessed 2026-11-13

Benchmarks are Edge0's own

Edge0 published six scores for Edge0-145B and said "Edge0, Qwen, DeepSeek, and Claude results come from Edge0's internal evaluations". The full evaluation setup, hardware and reproduction steps are promised with a technical report that is not yet linked. Treat every figure below as vendor-supplied.

Benchmark	Edge0-145B (vendor-reported)
DeepSWE 1.1	64.6
Terminal-Bench 2.1	86.52
HLE	35.4
GPQA-Diamond	90.4
IFBench	80.6
AA-LCR	79.3

The post compares against Qwen3.8-Flash-Next, Qwen3.8-27B, DeepSeek-V4-Flash-0731, Flint 40B, Muse Glimmer, Gemma 4 and Claude-Opus-4.6 (Max). It reports 1,100 or more tokens a second for prefill and 80 or more for decode.

What it needs to run

The headline memory figure is a peak, and the post gives it as 29GB in one place and 28GB in another. The card's measurement is a single run on an RTX 5090 D with 24GB of VRAM: 23.25 GiB of VRAM plus 38.27 GiB of host RAM, with the NVFP4 build at 47.79 GiB on disk. The card says these are readings, not minimum requirements.

Building from source targets NVIDIA GPUs with CUDA 13. On Hacker News, a commenter named PaulKeeble said the NVFP4 format means Nvidia Blackwell hardware only and that the repository was a directory of files, not a ready GGUF download. Another, Zetaphor, wrote "This is not open source, this is open weights".

The model card showed 14,134 downloads over the last month and 100 likes on November 13. Edge0 said the technical report is coming but gave no date.

An independent check of the kind these scores lack already exists for another open-weights release: Runway's Praxis-1 was ranked fifth on an NVIDIA leaderboard. Separately, six Chinese models read as Blackwell-trained in a seed fingerprint report.