Edge0-145B Open Weights Put a 145B Model on One GPU
Edge0 open-sourced a 145 billion parameter model on November 12 that runs on a single GPU with 16GB of VRAM. Here is what the download actually gives you, and where it falls short.

Edge0 open-sourced Edge0-145B and Edge0-120B on November 12, and the number that matters is not the parameter count. It is the hardware bill. Edge0-145B open weights run on a single NVIDIA GPU with 16GB of VRAM or more, which puts a 145 billion parameter model on a machine a lot of people already own. That is a different design bet from the frontier labs, and it is aimed squarely at developers who want to run agents locally instead of renting inference by the token.
The model is built for personal-computer agents rather than datacenter serving. Edge0 trains a latent sparse mixture-of-experts design with predictive expert loading, so the GPU loads the experts it needs instead of holding all 145 billion parameters in memory at once. Only 10.12 billion parameters are active per token. That is the trick that makes the whole thing fit, and it is a deliberate trade: a smaller active footprint in exchange for a model that can live on one card.
What the Edge0-145B open weights actually include
The release is three artifacts, not a weights-only drop. The checkpoints live on Hugging Face as Edge0-145B-q3k and Edge0-145B-NVFP4, alongside 120B variants, under an Apache 2.0 license. The serving stack is edge0-infer and edge0-strata on GitHub, and the v0.2.0 release notes dated November 11 cover Windows x64 and arm64 packages, plus an Anthropic Messages API on the local server. One detail worth knowing before you download anything: the 145B weights do not ship inside the zip. You fetch the checkpoints separately, and the minimum spec is CUDA 13 with compute capability 7.5 or better.
Edge0-145B takes text and images in and text out, and it is pitched at the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. That is the same job description Meta used for its tiny Muse Ember model this month, but from the opposite end. Meta made the model small. Edge0 made a big model fit on small hardware.
Why one GPU changes the math
Compare it with the other big open release of the month. Kimi K3.1 Pro is a 2.8 trillion parameter model with a 1M token context, and its files take up about 1.56 TB. It is downloadable, but it is not something you run on a laptop. It needs serious GPU hardware, the kind of setup that costs real money to rent or own. Edge0-145B is the mirror image: a smaller model that trades raw capability for the ability to run where you already are.
That matters for three kinds of users. First, developers building local agents that touch private data, where sending documents to an API is not an option. Second, researchers who want to inspect weights, fine-tune, and run experiments without worrying that an API will change underneath them. Third, anyone whose inference bill has quietly become a line item worth optimizing. A model that runs on hardware you own has a fixed cost, and fixed costs are easier to plan around than per-token pricing.
The frontier labs are not standing still either. Claude Fable 5.2 took the top spot on the Artificial Analysis Intelligence Index this month while cutting its price 20 percent. The gap between the best hosted models and the best downloadable ones is still wide. Edge0 does not close it. It just makes the downloadable side of the gap a lot more useful.
The honest limits
Two caveats before you get excited. First, most of the benchmark rows Edge0 published are its own internal evals, not independent results. Treat them as a claim, not a verdict. Second, 16GB of VRAM is a minimum, not a recommendation, and a 145B model will still be slow on a single consumer card compared with a hosted API. Local does not mean fast. It means private, fixed-cost, and under your control.
There is also the question of what 145B buys you that the smaller models do not. Edge0's own line began with a 35B phone-scale release in September. The 145B is a step up in capability, but the honest answer to whether it beats a good 70B model on your specific task is to run it. That is now cheap enough to do, which is the point.
Who should try it first
If you build agents, run sensitive documents through models, or just want to see what a 145B model feels like without a cloud bill, Edge0-145B is worth an afternoon. Download the checkpoints, spin up edge0-infer, and point your existing agent harness at the local server. The Anthropic Messages API compatibility means a lot of existing tooling will just work. If it is too slow on your card, the 120B variant is there too. The open weights era keeps getting more practical, and this release is the most practical one yet for a single GPU.


