Edge0-145B: An open 145-billion-parameter model that's meant to run on your laptop

What it's about

On November 12, 2026, the team behind the Edge0 model family open-sourced two large language models: Edge0-145B and Edge0-120B. They also released their own inference framework, which is the software that actually runs the model on your machine, along with installation tools. The blog post lists a contact address at autoark.ai, which points to the company AutoArk as the publisher.

The developer calls Edge0-145B the first open-source model with 145 billion parameters that runs on your own device rather than in a data center. It's supposed to keep up with leading cloud models at coding, reasoning and agentic tasks. Agentic means the model works through several steps on its own and uses tools along the way. According to the developer, it runs on a suitable laptop such as a MacBook Pro. Peak memory use is said to be 29 GB, with output above 80 tokens per second.

The weights are on Hugging Face, and the framework and guides are on GitHub. If you don't want to set anything up, you can use the Edge0 app. The developer says it works offline, without an account and without usage limits.

How a 145B model is supposed to fit into 29 GB

Normally that wouldn't be realistic. Even with 4-bit quantization, a model this size would take up roughly 80 to 90 GB. Quantization shrinks a model by storing its numbers at lower precision. Edge0 combines four techniques to fit on a laptop anyway:

• Architecture: Edge0-145B is a Mixture-of-Experts (MoE) model. It's made up of many specialized subnetworks, the "experts", and only a few of them work on any given token. Only about 10 billion parameters are active per step. What's new is that groups of 64 experts share two large matrices. Each expert keeps only a small set of its own parameters that captures what makes it different. Until now, every expert had a complete set of its own. The result is less data to store and move around.
• Training: The model was trained in this local-first design from the start. It isn't a cloud model that was compressed or pruned after the fact.
• Loading: Data is spread across three memory tiers, and reading and computing happen at the same time. Frequently used "hot experts" stay in VRAM (video memory). An expert library of about 40 GB sits in system RAM and gets loaded on demand. Around 48 GB of vocabulary data stays on the SSD and is only read when needed.
• Compute scheduling: The GPU and CPU work in parallel. On top of that there's speculative decoding: the system drafts several tokens at once and checks them together, so it only has to read the weights once for several tokens.

The developer says the model processes input at more than 1,100 tokens per second. This first step is called prefill. At that rate, dozens of pages of code should be read in quickly. Typical power draw is said to be about 120 watts, and the model reportedly doesn't throttle. The blog post doesn't clearly explain how the 29 GB peak memory fits with a 40 GB expert library in RAM. Presumably only part of it counts toward the measured peak, but that isn't spelled out.

Benchmarks: strong, but self-reported

The developer compares Edge0 with open models and with Claude Opus 4.6 (Max). All values are percentages, and higher is better:

DeepSWE and Terminal-Bench measure how well a model works as a coding agent, for example in the terminal or on real software projects. HLE ("Humanity's Last Exam") and GPQA-Diamond test hard expert knowledge and reasoning. IFBench measures how precisely a model follows instructions. AA-LCR tests comprehension of very long texts.

Edge0-145B's biggest lead is in coding. Compared with Qwen3.8-27B, which the developer uses as the usual reference point for local models, it scores 53.1 percent higher on DeepSWE and 18.5 percent higher on Terminal-Bench. On knowledge and reasoning the gap is small. On HLE and GPQA, Qwen3.8-Flash-Next and Claude actually come out ahead.

It matters where these numbers come from. The scores for Edge0, Qwen, DeepSeek and Claude come from Edge0's internal tests. The developer took the scores for Gemma 4 and other models from Artificial Analysis and the "Recursive Flint" evaluation. The test setup, hardware and reproduction instructions are supposed to be published with a technical report. There are no independent measurements yet.

Averaged across six benchmarks, Edge0-145B scores 72.8 points with 29 GB of measured memory. Qwen3.8-Flash-Next scores 71.7 points and, by the developer's estimate, needs about 70 GB. DeepSeek-V4-Flash scores 70.1 points at about 87 GB. The competitors' memory figures are only estimates for a typical 4-bit version and weren't measured.

What the model does in the demos

The developer shows several projects the model supposedly completed locally over many loops: planning, writing code, running it, fixing bugs and checking the result with screenshots.

• Explainer video: A video about how Tesla's Roadster might fly. The model reportedly handled research, script, graphics and editing in under an hour.
• 3D scene of Manhattan: The model built the scene in Blender from text and reference images and tweaked materials and lighting. According to the developer, a cloud model would have run up more than $150 in token costs for this.
• Medieval city: Also made in Blender, with city walls, more than a thousand houses and a Gothic cathedral, plus a 20-second camera flight. The model reportedly spotted and fixed errors such as objects clipping into each other on its own.
• Drone: A single sentence turned into the control program for a real drone. It flew through obstacles and found a red backpack.
• Circuit board design: The model designed a two-layer circuit board in KiCad, a PCB design program. It then rendered the manufacturing process as an animation in Unreal Engine 5.

The cost comparisons with cloud models and human teams are the developer's own claims. The blog post doesn't say how often runs like these succeed or how much human help was involved.

What you need to try it

The developer says Edge0-145B is supported on macOS, Windows and Linux from day one. That includes Apple M-series chips and Nvidia platforms such as the RTX 4090, RTX 5090 and RTX Spark. Peak memory is said to be 29 GB. The developer names laptops with 48 GB of memory as target devices, for example a MacBook Pro with that much unified memory, as well as workstations with 64 GB. An RTX 4090 has only 24 GB of VRAM, so there the model depends on the offloading to system RAM described above. How much RAM you'd need in that case isn't stated. With an expert library of around 40 GB, 64 GB of RAM is a reasonable assumption. Plan for SSD space for the weights, including around 48 GB of vocabulary data alone. Well over 100 GB in total is probably realistic. The developer doesn't give the exact download size.

Important: the memory figures only apply to Edge0's own inference framework. There's no evidence of similar memory use with standard tools like llama.cpp or Ollama, and it's unclear whether GGUF files exist. The easiest way to start is the Edge0 app. For smartphones there are smaller siblings: according to the developer, Edge0-35B ran on an iPhone 18 Pro Max with about 2.5 GB of memory at more than 37 tokens per second.

Background: the Edge0 model families

By its own account, Edge0 isn't new to this. In September it released Edge0-35B for smartphones, which hit number 1 on Hugging Face's trending list. In total, the team says it has released 200 models across four families:

• Edge0 for agents and coding
• Audio8 for speech recognition and text-to-speech
• Decision0 for fast decisions in about 40 milliseconds
• Embedding0 for local memory and RAG, meaning lookups in your own documents

The models have reportedly been downloaded more than 15 million times in three months, run on more than 5 million devices and have more than 110,000 customer organizations. Edge0 says that at a major headphone maker, 92 percent of AI usage runs directly on the device.

Our take

If the claims hold up, Edge0-145B is an important step for local AI. Until now, models for your own devices have mostly been smaller than 10 billion parameters. Apple's on-device model has about 3 billion, Google's Gemini Nano 1.8 or 3.25 billion. An open model on par with cloud models that runs on a laptop would be especially interesting for developers. Your code would stay on your own machine, and there'd be no ongoing API costs. The approach of shared expert matrices with data spread across GPU, RAM and SSD is technically exciting.

Still, caution is warranted. All the key numbers come from the developer: the benchmarks, the memory figures and the cost comparisons in the demos. The competitors' memory figures are only estimates. On broad knowledge (HLE), Claude Opus 4.6 stays ahead. And the memory savings depend on the developer's own framework. Independent tests still have to show how well it runs on ordinary hardware in practice. The blog post doesn't mention the license either.