Edge0-145B Puts a 145-Billion-Parameter Coding Agent on a Single Consumer GPU

Edge0 has released Edge0-145B and Edge0-120B, two open-weight language models built to run coding-agent work on a single consumer NVIDIA GPU. The 145-billion-parameter model uses a latent sparse mixture-of-experts design that keeps only about 10 billion parameters active per token, with weights on Hugging Face and the inference tools on GitHub — though the licence limits commercial use and the headline speed figures are the developer's own.

Nov 17, 2026 - 03:04
 0  5
Edge0-145B Puts a 145-Billion-Parameter Coding Agent on a Single Consumer GPU
A software developer in a denim shirt works at a desk with a laptop and an external monitor, both displaying lines of programming code with syntax highlighting.

Edge0-145B Puts a 145-Billion-Parameter Coding Agent on a Single Consumer GPU

Edge0 has released Edge0-145B, a 145-billion-parameter language model designed to run on a single consumer graphics card, alongside a smaller 120-billion-parameter sibling. The models arrived on November 12, 2026, with their weights published on Hugging Face and the software needed to run them on GitHub — a release aimed squarely at developers who want a capable coding agent running on their own machine instead of in a distant data centre.

It is the most ambitious step yet for a project that began with a 35-billion-parameter model for phones in September. This time the target is not a handset but a well-equipped desktop or laptop, and the pitch is that a model of this size no longer has to live behind a cloud API.

A 145-billion-parameter model that only wakes part of itself

The trick is in the architecture. Edge0-145B is a latent sparse mixture-of-experts model: it holds 145 billion parameters in total, but only about 10.12 billion of them are active for any single token. Across its 48 layers sit 512 experts each, of which the router picks 10 — plus one shared expert — per token.

That matters because local inference usually runs out of memory and memory bandwidth long before it runs out of computing power. Instead of loading the whole model into graphics memory, Edge0's engine keeps a small set of experts on the card and streams the rest from system memory as needed, a technique the project calls predictive expert loading. On an RTX 5090 D with 24 GB of video memory, the NVFP4 build used 23.25 GiB of VRAM plus 38.27 GiB of host RAM with that residency mode enabled — so "one GPU" is accurate, but it comes with a large helping of system memory attached.

The model is derived from Qwen3.8-Flash-Next, Alibaba's model, but Edge0 says it is not a simple relabelling or re-quantisation: the expert layer was rebuilt, with experts running at a narrow latent width, and the model was then post-trained on Edge0's own data. It was trained natively at 4 bits rather than compressed after the fact, and it supports up to 256K tokens of context. Two builds are published — Q3_K at 40.81 GiB and NVFP4 at 47.79 GiB.

What it can do, according to Edge0

Edge0 presents the model as an agent first and a chatbot second. Its public demonstrations lean on multi-step tool use — building scenes in Blender, editing circuit-board designs, writing small games — rather than single-shot questions.

On the numbers the project reports, the NVFP4 build reaches 98.3 tokens per second for agent decode on an RTX 5090 D, rising to 121.2 tokens per second with the Q3_K build. On benchmarks, Edge0 reports 64.6 on DeepSWE 1.1 and 86.5 on Terminal-Bench 2.1, ahead of the Qwen model it was derived from and, on those tests, competitive with leading cloud models.

Those figures should be read with care: every benchmark and speed number published so far comes from Edge0's own internal evaluations, and no independent measurements have appeared yet. The comparison with Claude Opus 4.6 in the project's materials is likewise Edge0's own work.

What you need to run it

The release is a stack of three parts, not a weights-only drop. Alongside the models, Edge0 published edge0-infer, its inference engine with a command-line interface, a local server compatible with OpenAI's API, and a desktop chat app, plus edge0-strata, the deployment tooling.

The hardware requirements are specific: Linux or Windows with an NVIDIA GPU, CUDA 13, compute capability 7.5 or newer, at least 16 GiB of video memory and 20 GiB of system memory. The Q3_K build also runs on Apple Silicon Macs through Metal, according to the model card, but the NVFP4 build is CUDA-only, and the Edge0 Chat desktop app's 0.2.0 release ships Windows packages only. The model weights are downloaded separately — they are not bundled with the app.

The licence is the catch

Edge0-145B is open-weight, not open-source, and the distinction carries real weight here. The weights sit under the Edge0 Community Model License 1.0, layered on top of the Qwen licence. Personal use, research, internal business use and commercial use of the model's outputs are permitted at no charge.

But selling or commercially bundling the model itself, operating a commercial model service, or offering an independent commercial coding or office assistant built on it all require a separate written licence from Edge0. The inference framework itself is Apache 2.0, so the restriction applies to the weights, not the tooling around them. Anyone planning to build a product on the model should read the licence first rather than assume the usual open-model freedoms.

Why it matters

A capable coding agent that runs locally changes the economics and the privacy picture at once: no per-token API bill, no code leaving the machine, and no account or usage quota. That is the same direction the wider local-AI movement has been heading — llama.cpp reached version 1.0 with a stable API for local apps in the same week — and Edge0's release pushes the size ceiling for what a single consumer card can plausibly host.

Whether the model lives up to its self-reported numbers in independent testing is the open question, and so is how many developers will accept the licence terms. For now, though, the release is a concrete step toward serious agent work on hardware many developers already own.

Sources

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Angry Angry 0
Sad Sad 0
Wow Wow 0
Nahid Joy I’m Nahid Hasan Joy, a technology writer, web developer, and digital enthusiast with a strong interest in the ever-changing world of technology.