Edge0-145B: A 145B AI Model That Runs on Your Laptop for Free

Until this week, a 145-billion-parameter AI model was a cloud job. You rented it by the token, sent your data to someone else's servers, and paid again every time you used it. Edge0-145B breaks that pattern, at least on paper: a 145B model you download for free and run on your own laptop, with no subscription and no per-token bill. The weights landed on Hugging Face on November 11, 2026, and the company's launch post followed on November 12. Within four days the two main builds had passed 61,000 downloads and the project's GitHub repository had passed 11,500 stars. This article checks what Edge0 actually released, what its own numbers show, what it costs in practice, and the license detail that matters before you build on it.

What Edge0-145B actually is

Edge0-145B is a latent sparse mixture-of-experts model with 145 billion total parameters, of which about 10.12 billion are active on each token. It has 48 layers and 512 experts per layer, with 10 routed experts plus one shared expert active at a time, and a 256,000-token context window. It was trained natively at 4-bit precision, which is the core of the design: instead of training a big cloud model and compressing it afterward, Edge0 built the on-device architecture into training from the start. The model is derived from Qwen3.8-Flash-Next, Alibaba's open-weight model, and that lineage shows up in the license terms below.

The release ships in two builds. Q3_K weighs 40.81 GiB and NVFP4 weighs 47.79 GiB. A smaller sibling, Edge0-120B, launched the same day with 120 billion total parameters and 8.64 billion active per token. Both run through Edge0's own inference framework, edge0-infer, with serve and chat commands, or through the Edge0 Chat desktop app. Day-one support covers macOS, Windows, and Linux, on Apple M-series chips and NVIDIA platforms including RTX Spark, RTX 4090, and RTX 5090. The framework code is under the Apache 2.0 license. The weights are not, which matters, and we cover that in the license section.

Edge0 calls the model the world's first open-source on-device AI model with 145 billion parameters. That is the company's own claim, and it rests on a specific definition of on-device: a model this size running on a personal computer rather than a server. Before September 2026, on-device models mostly stayed below 10 billion parameters. Apple's on-device foundation model runs at about 3 billion parameters, and Google's Gemini Nano came in 1.8B and 3.25B versions, according to Edge0's launch post. Edge0's own earlier phone model, Edge0-35B, arrived on September 10 and ran on phones at under 2.5GB of memory. The 145B release is the same idea scaled up to laptop hardware.

The laptop numbers, and what they leave out

The headline claim is about memory. Edge0 says the model runs on a compatible laptop with 29GB of peak memory, processes prompts at more than 1,100 tokens per second, generates more than 80 tokens per second, and draws about 120 watts of typical power, holding 128K to 256K of context at about the same speed. One passage in the company's own post says 28GB instead of 29GB, so treat the figure as about 29GB rather than exact. Either way, this is a job for a well-equipped laptop, not any laptop: the post describes a 48GB-class machine, and 29GB of peak memory leaves little room on a 32GB laptop once the operating system and other apps take their share.

The project's GitHub documentation gives measured speeds on specific machines, and those numbers are more useful than the laptop headline because they name the hardware:

Two things stand out. First, the gap between a desktop RTX 5090 and an Apple M2 Ultra is wide: 121.2 tokens per second of agent decode on the 5090 against 57.0 on the M2 Ultra, and the 16K prefill gap is wider still. Second, the last row is a constrained-budget test on the same RTX 5090 D with VRAM capped at 16GB and system RAM capped at 20GB, not a minimum requirement. It shows the model still runs when memory is tight, at 44.9 tokens per second, but that is less than half the speed of the uncapped card. The MacBook Pro figures in the launch post, more than 80 tokens per second, sit between the M2 Ultra measurement and the desktop cards, and Edge0 does not name the exact MacBook configuration behind them.

The design behind these numbers is worth a short explanation, because it is also the reason the model fits at all. Edge0 shares large matrices across groups of 64 experts, so each expert keeps only a small set of parameters that capture its differences. Frequently used experts stay in GPU memory, an expert library of about 40GB sits in system memory, and about 48GB of vocabulary data sits on the SSD and is read on demand. The GPU computes cached experts while the CPU handles the rest in parallel, and speculative decoding checks several candidate tokens at once. The result is a 145B model that needs roughly 29GB of peak memory instead of the roughly 70GB Edge0 estimates for a standard 4-bit deployment of its parent model, Qwen3.8-Flash-Next. That 70GB figure is Edge0's own estimate, not a measurement, and the company labels it as such.

Edge0's own benchmark table, read carefully

All benchmark figures in this section come from Edge0's own evaluations, published in its launch post and model documentation. No independent lab had reproduced them at launch, and the technical report with the full evaluation setup was still to come. Treat them as the company's claims about its own model, not as settled results.

Read the table the way Edge0 presents it, and the pattern is clear. Edge0-145B leads on the two coding and agent benchmarks, DeepSWE 1.1 at 64.6 and Terminal-Bench 2.1 at 86.52, ahead of its parent Qwen3.8-Flash-Next at 58.7 and 83.3, DeepSeek-V4-Flash-0731 at 54.4 and 82.7, and Claude Opus 4.6 (Max) at 78.2 on Terminal-Bench. Claude has no public DeepSWE 1.1 score in the table, so that column has no cloud comparison. The picture flips on knowledge and reasoning tests. Claude leads Humanity's Last Exam at 40.0 against Edge0's 35.4, and Qwen3.8-Flash-Next leads GPQA-Diamond at 91.7 against Edge0's 90.4 and IFBench at 81.3 against 80.6. Edge0's mean score across the six benchmarks is 72.8, against 71.7 for Qwen3.8-Flash-Next and 70.1 for DeepSeek-V4-Flash-0731, on Edge0's own arithmetic.

Two comparisons need context. First, the Terminal-Bench lead over Claude is the strongest single result in the table, but it is a single benchmark run by the company that built the model, and Claude's IFBench score of 62.5 is far below the rest of the field, which suggests the harness or prompting favors some models over others. Second, the table includes Flint 40B at 54.9 on Terminal-Bench 2.1, from the Recursive Flint evaluation we covered in October, which puts Edge0's 86.52 in perspective against another laptop-class coding model. The honest summary is narrower than the launch headline: Edge0-145B looks strong on coding and agent work by its own tests, and ordinary on knowledge tests, with no independent confirmation yet.

What it costs: $0, plus a GPU you already own

The model itself costs nothing. The weights are free to download from Hugging Face, there is no subscription tier, and there is no hosted API to bill you per token, because Edge0 does not run one. The cost of using Edge0-145B is the hardware it runs on and the electricity it draws, which is the whole point of the release and also its real price tag.

That framing only saves money if you already own suitable hardware or were going to buy it anyway. A laptop with 48GB of memory or a desktop with an RTX 4090 or 5090 is a large purchase, and nobody should buy one for a single model on launch-week benchmarks. The saving is real for a different group: developers and teams who already have the hardware and currently pay a cloud provider per token for coding and agent work. For them, every local run replaces a metered API call, and long agent sessions with dozens of planning, execution, and review rounds are exactly the workloads where cloud token bills grow fastest. Edge0's own demos lean on this point. Its Manhattan city-scene demo, built in Blender in under an hour, carries Edge0's estimate that a leading cloud model would have cost more than $150 (about £113, about C$214) in tokens for the same job, and that manual work would have cost $12,000–18,000 (about £9,079–13,619, about C$17,110–25,664). Those are Edge0's estimates, not measured bills, and the currency figures use November 15, 2026 market rates of 1 USD = 0.7566 GBP and 1.4258 CAD, verified at publication. Treat the comparison as the company's marketing math, not as a price you can verify.

The license catch

This is the part of the release that needs the most care, because the launch post's "open-source" wording is simpler than the actual terms. The inference framework code is under the Apache 2.0 license. The model weights are under the Edge0 Community Model License 1.0, a separate license that applies to Edge0's own model contributions. And because the model is derived from Qwen3.8-Flash-Next, Qwen materials remain governed by the Qwen Community License 1.0, which Edge0 reproduces unchanged. Edge0 states plainly that its permission does not waive Qwen's attribution or separate commercial authorization requirements.

In practice, the Edge0 license permits personal use, research, education, modification, noncommercial sharing, internal use, and commercial use of generated outputs, and it does not require you to disclose private modifications or show an Edge0 interface badge. What it does not give you is a plain Apache-style grant over the weights, and the Qwen upstream terms add their own conditions. If you plan to ship a product built on Edge0-145B, read the Edge0 Community Model License and the Qwen Community License in the repository before you commit, and check whether your use needs a separate commercial authorization. For personal use, research, and internal tools, the terms are permissive. For resale or hosted inference, they are not a blank check. That distinction is the difference between this release and a model under Apache 2.0 or MIT, and it is the single most important detail in the release for anyone building a business on it.

How to run it

Edge0 ships the model and the inference engine together, and the setup follows the project's GitHub documentation. The short version, from the repository's getting-started guide:

1. Clone the repository with git clone https://github.com/Edge0-AI/Edge0.git edge0 and build the edge0-strata engine for your platform: Linux or Windows with an NVIDIA GPU and CUDA 13, or macOS on Apple Silicon with Metal.
2. Download a model folder from Hugging Face, either Edge0-145B-q3k at 40.81 GiB or Edge0-145B-NVFP4 at 47.79 GiB. Keep the folder as it is; the engine identifies the model from the model.json file inside it.
3. Run it with edge0-infer serve for a local server or edge0-infer chat for a terminal chat session. If you want no build step, the Edge0 Chat desktop app runs the same models with no account and no usage quota, and its core capabilities work offline.
4. Pick the build for your hardware. Q3_K is the smaller download and the faster decode on NVIDIA cards in Edge0's measurements, while NVFP4 targets Blackwell GPUs. On Apple Silicon, Q3_K is the supported build.

Budget for disk space beyond the weights. The engine keeps about 48GB of vocabulary data on the SSD and reads it on demand, so a machine with a small drive will feel the download twice. The project's documentation also includes memory-saving options for constrained setups, including the capped 16GB VRAM configuration in the speed table above, but those trade speed for fit. If your first run is slow, check whether the model folder is complete and whether your GPU is actually being used before assuming the model is at fault.

What Edge0 says it can do

The launch post includes five demonstrations, all produced by Edge0 itself with no independent verification. They are worth describing because they show what the company thinks the model is for, and they should be read as demos, not as reviews.

The coding demo is the most grounded. Edge0-145B plans a task, writes code, runs it, debugs errors, and checks results visually, iterating until the work is done. The benchmark table above supports the coding claim better than the other demos do, since DeepSWE and Terminal-Bench measure exactly this kind of agent work. The video demo produced an explainer about a flying Tesla Roadster, covering research, scriptwriting, asset generation, and editing in under an hour, all locally. The 3D demo built a Manhattan city scene in Blender in under an hour, then a medieval capital with more than a thousand homes and a 20-second fly-through, inspecting renders and correcting its own work as it went. The hardware demo wrote flight-control and visual-recognition code for a real drone from a single instruction and flew it indoors through obstacle poles to find a red backpack. The PCB demo took a board from component placement in KiCad through routing to an assembly animation in Unreal Engine 5.

None of these demos has an independent reproduction yet, and the time and cost comparisons attached to them are Edge0's own. The 3D and video demos are the ones to be most careful with: a model that writes Blender scripts and inspects renders is doing real agent work, but the quality of the output is a matter of taste and the demos show Edge0's chosen takes. The drone and PCB demos are the most interesting technically, because they involve real hardware and real design tools rather than generated media. If you test the model yourself, start with coding and agent tasks, where the benchmarks give you something to compare against, before trusting it with hardware or design work.

Who should try it, and who should wait

Try it this week if you already own a 48GB laptop or a desktop with an RTX 4090, 5090, or RTX Spark GPU, and you pay a cloud provider for coding or agent work today. Download the Q3_K build, run your real tasks through it, and compare quality, speed, and electricity cost against your current setup. The download is free, the app needs no account, and a direct comparison on your own work is worth more than any benchmark table, including Edge0's.

Wait if you would need to buy hardware for it, or if you need a hosted API, a service-level agreement, or a support contract, because none of those exist. Wait as well if your work depends on current information or cross-device memory, since the model runs locally and has no live web access of its own; Edge0 says features that depend on online information still need a connection. And wait if you need settled licensing for a commercial product, until you have read both the Edge0 and Qwen license texts and confirmed your use is covered.

The broader context is a shift Edge0 describes in three stages: cloud-first AI before September 2026, everyday tasks local and complex tasks in the cloud after its Edge0-35B phone release, and now complex tasks local with cloud support. That framing is the company's own, and the 145B release is a real step toward it for coding work. Whether it holds up for the harder demos depends on independent testing that has not happened yet.

Common mistakes

Treating "open-source" as Apache 2.0: the framework is Apache 2.0. The weights are under the Edge0 Community Model License 1.0, with Qwen Community License terms on the upstream material. Read both before building a product on the model.

Downloading the wrong build for your GPU: Q3_K and NVFP4 are separate 41GB and 48GB downloads with different speed profiles. Q3_K is faster on the cards Edge0 measured, and NVFP4 targets Blackwell GPUs. On Apple Silicon, use Q3_K.

Judging it by the mean benchmark score: Edge0's 72.8 mean across six benchmarks hides a split: it leads on coding and agent tests and trails on knowledge tests. Pick the benchmark that matches your work, not the average.

Assuming your laptop qualifies: the 29GB peak memory figure is for a compatible, well-equipped laptop. A 32GB machine with a full desktop session running will struggle, and a phone cannot run this model at all. Edge0-35B is the phone model; Edge0-145B is the laptop one.

Expecting a hosted API: there is no Edge0 API to call, no dashboard, and no usage-based billing. Local inference is the product. If your application needs a cloud endpoint, this release does not provide one.

Frequently asked questions

Is Edge0-145B really free?

Yes, the weights are free to download and there is no subscription or per-token charge, because there is no hosted service. You still pay for the hardware it runs on and the electricity it uses, and Edge0 reports about 120 watts of typical power on its laptop setup.

Can Edge0-145B run on a MacBook?

On a compatible, well-equipped MacBook, yes. Edge0 reports 29GB of peak memory, more than 1,100 tokens per second of prefill, and more than 80 tokens per second of decode on a MacBook Pro, though it does not name the exact configuration. Apple Silicon support uses the Q3_K build through Metal. A base-model MacBook with 16GB or 24GB of memory will not fit it.

How is Edge0-145B different from Edge0-35B?

Edge0-35B is the phone model, released September 10, 2026, running at under 2.5GB of memory with 37 or more tokens per second on an iPhone 18 Pro Max, according to Edge0. Edge0-145B is the laptop model, with about four times the parameters, aimed at coding and complex agent work that the phone model cannot take on.

Is Edge0-145B better than Claude Opus 4.6?

On Edge0's own tests, it scores higher on Terminal-Bench 2.1, 86.52 against 78.2, and has no DeepSWE comparison because Claude has no public score there. On Humanity's Last Exam, Claude leads 40.0 to 35.4, and on GPQA-Diamond Claude leads 91.3 to 90.4. The honest answer is that it looks stronger on coding and agent work and weaker on knowledge tests, with no independent confirmation yet.

Can I use Edge0-145B in a commercial product?

Personal use, research, education, and commercial use of generated outputs are permitted under the Edge0 Community Model License 1.0, but the weights are not under Apache 2.0 or MIT, and Qwen Community License terms apply to the upstream material, including attribution and separate commercial authorization requirements. Read both license texts in the repository before shipping a product, and check whether your use case needs an additional agreement.

Where can I download Edge0-145B?

The weights are on Hugging Face as Edge0-145B-q3k and Edge0-145B-NVFP4, and the inference framework and deployment guides are in the Edge0 repository on GitHub. The Edge0 Chat desktop app is available from the same repository and needs no account.

Sources

• Edge0 — Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer (November 12, 2026)
• Edge0 — GitHub repository, README and model documentation (Edge0-AI/Edge0)
• Hugging Face — Edge0/Edge0-145B-NVFP4 model card
• Hugging Face — Edge0/Edge0-145B-q3k model card
• arXiv — Edge0 technical paper (2609.18063)
• DEV Community — Edge0 AI open-sources Edge0-145B (November 14, 2026)
• AI Core Updates — Recursive Flint: Free 40B Laptop Coding Model Built Largely by AI
• AI Core Updates — Meta's Muse Spark Open: 816B AI Model You Can Download Free
• AI Core Updates — DeepSeek V4.1 Flash 1017: The $0.15 Coding Model and Its New Search API

Exchange rates used: 1 USD = 0.7566 GBP and 1.4258 CAD, November 15, 2026 market rates, verified at publication. Currency conversions are approximate. Benchmark and speed figures are Edge0's own reported results unless stated otherwise.