Open-weight models, deployment and licensing

Edge0 releases 145B and 120B latent-MoE weights for single-GPU inference

Four public GGUF/NVFP4 repositories target NVIDIA GPUs, but host-memory offload and custom model licenses make the practical deployment story more nuanced.

Open Model Weights published15 Nov 2026
Primary sourceEdge0 — official Edge0-145B Q3_K model card, weights and license
Source published2026-11-13
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original cinematic glossy CGI of two chrome neural accelerator towers, illuminated latent expert cubes and teal offload trails, representing Edge0 145B and 120B models.
Original high-gloss conceptual artwork of streaming mixture-of-experts inference; not a model manufacturer photograph, a hardware test or evidence of published performance. AI-generated editorial illustration; not a documentary photograph.

Edge0 has published 145-billion- and 120-billion-parameter latent mixture-of-experts model variants, extending its local inference effort beyond its earlier 35B and 8B releases. The project's November 13 announcement links four checkpoint repositories: Q3_K and NVFP4 for each model. These are derivatives of Qwen3.8-Flash-Next, not independently trained foundation models presented with a completely new license. The premise is that sparsely activated experts and a dedicated streaming runtime can make unusually large model weight collections usable on consumer-class NVIDIA hardware. Primary source ↗

The official 145B card lists 10.12 billion active parameters per token, 48 layers, a 256K context configuration and a 40.81 GiB Q3_K weight package. Edge0 reports 121.2 agent-decode tokens per second on an RTX 5090 D 24 GB. The 120B Q3_K file is listed at 27.98 GiB. These are publisher measurements and configuration claims, not Open Model Weights reproductions. A 24 GB GPU does not hold the entire larger weight set by itself: the project's residency and constrained-memory tests use system RAM and streamed expert placement.

Deployment also has a software constraint. For the new 145B and 120B releases, Edge0's supported build instructions point to Linux or Windows, an NVIDIA GPU, CUDA 13 and the edge0-strata inference engine invoked through edge0-infer. Older Edge0 models use a different native engine, and an Apple Metal throughput row on the card is not itself an announcement of macOS release support. Users should follow the specific model folder and engine documentation rather than infer compatibility from a generic model-hosting download button.

The licensing distinction is particularly important. The published weights carry Edge0 Community Model License 1.0 alongside applicable upstream Qwen conditions, while the framework code has different open-source terms. Personal research and internal use may be permitted, but commercial model hosting, reselling or independent coding assistants can require a separate agreement. Open Model Weights should distinguish publicly downloadable files from commercially unrestricted weights, pin exact revisions and leave vendor rankings attributed until they can be independently reproduced.

OMW REGISTRY WATCH

What this changes for the evidence layer

Treat 145B and 120B as distinct derivative checkpoints to investigate, not as automatically verified general-purpose open models. The Q3_K files are about 40.81 and 27.98 GiB respectively, and the official single-GPU route still relies on host-memory residency and a custom edge0-strata engine. Edge0 Community Model License 1.0 and upstream Qwen terms are not a blanket Apache/MIT grant; commercial serving and redistribution can need separate permission. Confirm public file manifests, checkpoint revisions and exact applicable licenses before a registry entry.

PRIMARY SOURCE

Edge0 — official Edge0-145B Q3_K model card, weights and license

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗