Edge0 open-sourced its 145B-parameter model on 12 November 2026 — weights on Hugging Face on 11 November, GitHub v0.2.0 the same day, and the announcement blog on 12 November. Edge0-145B is a latent sparse mixture-of-experts model: 145 billion parameters in total, with only about 10.12 billion active per token, across 48 layers and 512 experts per layer. It offers a 256K-token context window and native 4-bit weights, and it is derived from Alibaba's Qwen3.8-Flash-Next.
The headline numbers look strong: Edge0 claims a 29GB peak footprint, 80+ tokens per second decode, 86.52 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1. But these are Edge0's own claims with no independent verification — and 4Geeks notes that the first independent test, on an RTX 3090, tells "a more mixed story". Treat the benchmarks as the vendor's figures until more independent runs appear.
The part that matters most for anyone planning to use it is the licence. The code is Apache 2.0, but the weights are under the Edge0 Community Model License 1.0 — not Apache, not MIT. Commercial use of the model's outputs is allowed, but selling the model itself or running it as a commercial service needs separate permission from Edge0. On hardware, Linux and Windows setups need an NVIDIA GPU with CUDA 13 and 16GB or more of VRAM.
Key facts
| Item | Details |
|---|---|
| Model | Edge0-145B, latent sparse MoE |
| Open-sourced | 12 November 2026 (weights on Hugging Face 11 Nov; GitHub v0.2.0 11 Nov) |
| Parameters | 145B total, ~10.12B active per token |
| Architecture | 48 layers, 512 experts per layer, 256K context, native 4-bit |
| Derived from | Alibaba's Qwen3.8-Flash-Next |
| Code licence | Apache 2.0 |
| Weights licence | Edge0 Community Model License 1.0 (NOT Apache/MIT) |
| Claimed performance | 29GB peak, 80+ tok/s decode, Terminal-Bench 2.1 86.52, DeepSWE 1.1 64.6 — Edge0's own claims |
| Independent test | 4Geeks: first independent test (RTX 3090) tells "a more mixed story" |
| Hardware (Linux/Windows) | NVIDIA GPU + CUDA 13 + 16GB+ VRAM |
What this means
Edge0-145B is genuinely usable by individuals: a 145B model that activates only ~10B parameters per token is designed to run on a single strong consumer GPU, and the native 4-bit weights keep the footprint down. That is the appeal.
The catch is the licence split. "Open weights" here does not mean "do anything with it" — the weights licence allows commercial use of outputs, but selling the model or offering it as a commercial service needs Edge0's separate permission. Anyone building a product on top of it should read the Edge0 Community Model License 1.0 before assuming the Apache 2.0 code licence covers the weights too. And on performance, wait for independent benchmarks before trusting the headline numbers.
A note on verification
The release dates, architecture figures and licence terms are as reported for the 12 November 2026 open-weights release. The performance figures are Edge0's own claims, with no independent verification; the 4Geeks note about the first independent test (RTX 3090) is carried as reported.
FAQs
Is Edge0-145B open source?
Partly. The code is Apache 2.0, but the weights are under the Edge0 Community Model License 1.0 — not Apache or MIT — with limits on selling the model or running it as a commercial service.
How big is the model?
145 billion parameters in total, with about 10.12 billion active per token, a 256K-token context window and native 4-bit weights.
What hardware does it need?
On Linux or Windows, an NVIDIA GPU with CUDA 13 and 16GB or more of VRAM.
Are the benchmark scores verified?
No. The 86.52 Terminal-Bench 2.1 and 64.6 DeepSWE 1.1 figures are Edge0's own claims; 4Geeks notes the first independent test tells "a more mixed story".
Source: Edge0 open-weights release, 11–12 November 2026 (Hugging Face, GitHub, Edge0 blog); 4Geeks independent-test note. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!