Skip to content
Rush Commerce
Tools & Teardowns3 min read

Edge0-145B local AI model: 145B params, read the license

Edge0-145B runs a 145B coding model on one consumer GPU. The memory math and the license both have fine print. Here is what a small business can use it for.

Edge0 released Edge0-145B, a 145-billion-parameter local AI model built for coding and agent work, and calls it "open-source on-device AI." The engineering is real: about 10B parameters active per token, 4-bit native training, and published decode speeds near 80 tokens a second on one RTX 5090. Two things in the fine print matter more than the headline: how much memory it actually uses, and what the license lets you do with it.

What actually happened

On November 12, Edge0 open-sourced Edge0-145B and Edge0-120B with its inference framework. The company reports 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, 128K–256K context, and says the model fits on a 48GB laptop. It runs 512 experts per layer and routes 10 per token.

The Hugging Face model card gives the measured setup. On an RTX 5090 D with 24GB, the NVFP4 build used 23.25 GiB of VRAM plus 38.27 GiB of system RAM, with 79.6 tok/s chat decode. The card also lists an Apple M2 Ultra at 60.9 tok/s, and a capped 16 GiB VRAM + 20 GiB RAM setup at 36.8 tok/s.

The benchmarks are Edge0's own: 64.6 on DeepSWE 1.1 and 86.52 on Terminal-Bench 2.1. Edge0 ran the comparison models through its own evaluations too. No independent results yet.

The card says Edge0 rebuilt the expert layer of Qwen3.8-Flash-Next, then post-trained it. That is why there are two licenses. Edge0's parts use the Edge0 Community Model License 1.0. Qwen materials stay under the Qwen Community License 1.0. The Edge0 license permits internal use, research, modification and commercial use of outputs. Selling or bundling the model, running a commercial model service, or offering a commercial coding or office assistant built on it needs a separate written license. Only the framework code is Apache-2.0.

Why a local AI model matters for your business

For internal work, this is a real option. A developer workstation with a 24GB card and 64GB of RAM can run a strong coding model with no per-token bill, and your code never leaves the building. Overnight jobs that burn API tokens are the obvious first test: refactors, test generation, ticket triage.

But "open-source" in the announcement does not mean you can build a product on it. If you plan to ship an assistant to customers on these weights, you need Edge0's written permission and Qwen's terms. Read both before you build.

And spec the machine from the card, not the press release. Count VRAM plus system RAM, not "29GB."

Key takeaways

  • Edge0-145B and Edge0-120B shipped November 12 with an Apache-2.0 inference framework
  • The model card shows 23.25 GiB VRAM plus 38.27 GiB system RAM on an RTX 5090 D, about 80 tok/s chat decode
  • All benchmark scores are Edge0's internal evaluations, including the comparison models
  • The weights license allows internal use; a commercial model service or bundled product needs a separate Edge0 license, and Qwen terms also apply
  • Test it on internal batch and coding jobs first, where no per-token bill is the win

Local models cut the token bill only when the license and the hardware fit the job. We test open-weight models on your actual workload, read the license terms with you, and move the internal jobs that make sense off the API meter. Estimate what your AI work costs now, or tell us which jobs you want to run locally.

Sources: Edge0 announcement, Edge0-145B model card on Hugging Face.

  • #local-inference
  • #open-weights
  • #edge0
  • #coding-agents
  • #model-licensing
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.