Skip to content
Get App

Edge0-145B: A Free 145B AI Model That Runs Offline on Your Laptop (The Catch)

Breakdown of Edge Zero 145B, a free open-weight 145B MoE model claimed to run offline on a laptop, with benchmarks, demos, and hardware caveats.

Key Takeaways

  • Edge Zero 145B is a promising free local model, but its headline claims are company-reported and need independent testing.
  • Hardware requirements are the main limitation: the big models currently need a high-end NVIDIA GPU with CUDA 13, not a typical MacBook.
  • Strong candidate for privacy-sensitive or high-volume token workloads if your machine qualifies; check licensing before commercial use.

What the video covers

  • Edge Zero 145B is presented as a free, open-weight 145-billion-parameter model that can run offline on a personal laptop without cloud access or token fees.
  • The company claims about 29 GB peak memory, prompt processing above 1,100 tokens/s, and generation above 80 tokens/s, drawing roughly 120 W.
  • Architecture tricks cited: a mixture-of-experts design using only about 10B parameters per token, tiered loading of experts across GPU, system RAM, and SSD, and CPU/GPU scheduling with speculative decoding.
  • Benchmarks reported by the company show Edge Zero leading open models on Deep SWE (64.6) and Terminal Bench 2.1 (86.5), though Claude Opus 4.6 still wins on HLE.
  • All benchmark comparisons come from the company's own internal evaluations, so independent verification is still pending.
  • Published demos include offline chat, a coding agent building a game, a Blender 3D city scene, a circuit board design in KiCad, and drone flight code.
  • The catch: the GitHub README lists Linux or Windows with an NVIDIA GPU and CUDA 13 for the big models, with no Mac support for the 145B variant at the time of recording.
  • The model is built on Qwen 3.8 Flash Next and trained natively at 4 bits, and the weights are about 113 GB.
  • Licensing: the engine is Apache 2.0, but model weights use a custom community license that permits personal, research, and internal business use but requires a separate commercial license for selling model access or paid assistants.
  • The presenter recommends smaller Edge Zero models for users without suitable hardware and suggests testing on your own tasks rather than trusting demos.

Answers

Questions about this video

Can Edge Zero 145B run on a MacBook?

According to the GitHub README at the time of recording, the 145B and 120B models target Linux or Windows with an NVIDIA GPU and CUDA 13, and macOS is not listed for these sizes. The blog's MacBook example does not fully match the repo instructions, so Mac users should check the repository before downloading.

Are the benchmark results independently verified?

No. The comparisons, including the Claude Opus 4.6 numbers, come from Edge Zero's own internal evaluations, and the full reproduction setup is promised with a later technical report. Independent testers have not yet confirmed them.

Is Edge Zero 145B free to use commercially?

The inference engine is Apache 2.0, but the model weights use a custom community license that allows personal, research, and internal business use and selling outputs. Selling access to the model or building a paid coding assistant on it requires a separate commercial license.

Full Transcript — Download SRT & Markdown

00:00
Speaker A
This is a 145 billion parameter AI model running offline on a laptop, no cloud, no account, and no token bill. It's called Edge Zero 145B and the company just released the weights for free. Edge Zero says it needs about 29 gigabytes of
00:20
Speaker A
memory at its peak and writes more than 80 tokens per second. That's faster than you can read. And on one coding agent test, Terminal Bench 2.1, Edge Zero reports a score higher than Claude Opus 4.6 from a model on your own machine.
00:37
Speaker A
But there's a catch and most posts about this model skip it. The GitHub page tells a slightly different story about what hardware you actually need. I'll show you all of it with the real pages so you can decide if it's worth your
00:51
Speaker A
time. This is the full breakdown of Edge Zero 145B, what it is, how a model this big fits on a personal computer, the benchmarks, the real demos, the catch, and how to try it. If you like free AI tools explained
01:07
Speaker A
honestly, subscribe to the channel. It really helps. On November 12th, a company called Edge Zero published this announcement introducing Edge Zero 145B, frontier AI capabilities on your own computer. They're open sourcing two models, Edge Zero 145B and Edge Zero 120B together with their
01:30
Speaker A
inference framework and deployment tools. And they call it the world's first open source on device model with 145 billion parameters. The weights are already on Hugging Face. This is the model page for the 145B version. You can see the tags here, mixture of experts,
01:49
Speaker A
4-bit precision, agent, and coding. And people are grabbing it fast. The page shows more than 14,000 downloads in its first days. The engine that it is on GitHub. Edge Zero describes it as an open-source streaming mixture of experts inference framework.
02:08
Speaker A
Why does this matter? Because when a model runs on your own machine, you don't pay per token. Your code and documents never leave your computer, and it keeps working with no internet at all. Until now, local models were small,
02:22
Speaker A
and small models struggle with hard work. Edge Zero is trying to break that tradeoff. So how do you fit a 145 billion parameter model on a laptop?
02:33
Speaker A
Here are Edge Zero's own numbers. Peak memory, 29 gigabytes. They say it runs on a compatible laptop like a MacBook Pro without a server.
02:43
Speaker A
Prompt [snorts] processing above 1,100 tokens per second. That means it can read dozens of pages of code before it starts answering. And generation above 80 tokens per second. It draws about 120 watts and supports a context of 128,000
03:00
Speaker A
to 256,000 tokens at almost the same speed on long inputs. The first trick is in the architecture.
03:09
Speaker A
It's a mixture of experts. Of the 145 billion parameters, only about 10 billion are actually used for each token. And groups of 64 experts share two large matrices, so far less data has to be stored and moved. The second trick
03:28
Speaker A
is how it loads. The experts it uses most stay in GPU memory. A library of about 40 gigabytes of experts sits in normal system memory, and about 48 gigabytes of vocabulary data stays on the SSD and is read only when needed.
03:45
Speaker A
And the third trick is scheduling. The GPU and CPU work at the same time, and a technique called speculative decoding checks several possible tokens at once.
03:56
Speaker A
Edge Zero also says the model was trained for this on device design from day one, not squeezed down after training. A simple way to picture it, imagine a giant library. You don't need every book on your desk. You keep the
04:11
Speaker A
few you're using right now on the desk, the rest on nearby shelves, and the rarely used ones in the basement. That's exactly what Edge Zero does with the model's experts. Now the benchmarks.
04:23
Speaker A
Here's Edge Zero's table. Edge Zero 145B scores 64.6 on Deep SWE, a real software engineering test, and 86.5 on Terminal Bench 2.1. For comparison, Qwen 3.8 Flash Next scores 58.7 on Deep SWE, and DeepSeek V4 Flash scores 54.4.
04:48
Speaker A
So on coding, Edge Zero leads these open models. And here's the comparison everyone is sharing. Claude Opus 4.6 at max settings scores 78.2 on Terminal Bench in this table, below Edge Zero's 86.5, but Claude still wins on the very hard
05:07
Speaker A
HLE test, 40 versus 35.4. Here's that test as a simple chart. Edge Zero on top, then Qwen, DeepSeek, and Claude Opus 4.6. But read the small print under the table. Edge Zero, Qwen, DeepSeek, and Claude results all come
05:27
Speaker A
from Edge Zero's own internal evaluations. So a fair warning, these are the company's own numbers. Edge Zero says the full evaluation setup and reproduction instructions are coming with its technical report.
05:41
Speaker A
Until independent testers run it, treat the Claude comparison as a claim, not a fact. The more interesting chart might be this one.
05:50
Speaker A
Edge Zero combines a mean score of 72.8 across six benchmarks with 29 GB of measured peak memory. Qwen 3.8 Flash Next averages 71.7, but Edge Zero estimates it needs about 70 gigabytes in a standard four bit setup. DeepSeek about 87. So similar
06:13
Speaker A
quality in well under half the memory. Those competitor memory numbers are estimates though. Quick pause. If this is useful, subscribe. Next, the real demos Edge Zero published and then the catch. This is Edge Zero's official showcase. First, chat encoding with
06:32
Speaker A
Wi-Fi switched off. Look at the corner. It says offline on device and the live speed counter shows around 75 to 80 tokens per second. Here it's working as a coding agent inside a terminal building a small block game. Edge Zero
06:48
Speaker A
says the model plans the task, writes code, runs it, fixes failures, and checks the result with screenshots.
06:57
Speaker A
Next, video production. Edge Zero says the model generated this explainer about the Tesla Roadster's suction fan idea within an hour fully locally.
07:07
Speaker A
Then 3D. This Manhattan city scene was built in Blender by the agent within an hour according to Edge Zero. They estimate the same scene would take a person about a month by hand. And here's a medieval cathedral built from scratch
07:23
Speaker A
in Blender with a camera fly through inside. Finally, hardware. The model designed a two layer circuit board in KiCad and then rendered an assembly animation in Unreal Engine 5. Edge Zero also showed it writing flight code for a
07:39
Speaker A
real drone. These demos show what an agent can do with real tools, but remember they're chosen by the company.
07:47
Speaker A
The real test is your own tasks on your own machine, which brings us to the catch. Here's the GitHub page, the platform table.
07:56
Speaker A
For the 145B and 120B models, the engine is called Edge Zero Strata and it lists Linux and Windows with an Nvidia GPU and CUDA 13. Now, look at macOS, iOS, and Android for the big models. Just a dash.
08:14
Speaker A
At the time I'm recording, the README doesn't list Mac support for 145B. The smaller 35B model is the one that runs on Mac and phones. And the model section says it again. 145B and 120B run through Edge Zero Infer Serve and Edge Zero
08:32
Speaker A
Infer Chat on Linux or Windows with an NVIDIA GPU and CUDA 13. Their reference speed is on an RTX 5090 laptop GPU with 24 GB, 121 tokens per second for agent decoding.
08:49
Speaker A
So the blog's MacBook example and the GitHub instructions don't fully match yet. If you're on a Mac, check the repo before you download anything. Two more honest details. The README says these models are derived from Quen 3.8 Flash
09:04
Speaker A
Next and trained natively at 4 bits. So it's built on top of Quen, not from scratch. And the license. The engine is Apache 2.0, but the model weights use a custom Edge Zero community license. It's free for personal use, research, and
09:21
Speaker A
internal business use. And you can sell what you make with it, but selling access to the model or building a paid coding assistant on it needs a separate commercial license. So before you download, three practical points. The model files are about 113
09:38
Speaker A
GB. The big models currently need an NVIDIA GPU with CUDA 13. And Edge Zero talks about 48 GB laptops as the target.
09:49
Speaker A
If you don't have that hardware, start with the smaller Edge Zero models. If your machine qualifies, here's how to start, straight from Edge Zero's announcement. The model weights are on Hugging Face. The inference framework and deployment guides are on GitHub, and
10:04
Speaker A
there's an Edge Zero app that needs no setup, works offline with no account and no usage quota. In short, three steps.
10:13
Speaker A
One, check that your GPU and memory match the Readme. Two, install the Edge Zero app or the Edge Zero Infer engine.
10:22
Speaker A
Three, load the 145B or 120B folder and start chatting or serving it. What should you use it for first? A coding agent on your own private codebase.
10:34
Speaker A
Questions about documents you'd never upload to a cloud service and long files like big logs or contracts where a cloud model would cost real money every single time. And to be fair, Edge Zero itself says where the cloud still wins.
10:51
Speaker A
Very low power devices can't run 145B and tasks that need long-term memory across devices or very current information from the web still need cloud services. So, the verdict. A 145 billion parameter model running locally is real and the weights are free to
11:11
Speaker A
download. The claim that it beats Claude Opus 4.6 is the company's own test, not independent yet. And for the big model, you'll want an Nvidia machine for now.
11:22
Speaker A
If you have a strong Nvidia laptop or desktop and you care about privacy or token costs, this is one of the most interesting free models to test this month. If you're on a Mac, keep an eye on the GitHub repo for updates. If this
11:35
Speaker A
helped, subscribe and turn on notifications. I'll cover the independent test results as soon as they're out. Now tell me in the comments, would you run a model like this on your own machine or stick with the cloud?
Topics:Edge Zero 145Blocal AI modeloffline LLMmixture of expertsopen-source AIon-device inferenceQwenAI benchmarksNVIDIA CUDAfree AI tools

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

Or transcribe another YouTube video here →