Edge0-145B: A 145B Open Model That Runs on a Laptop

Updated: November 20265 min readBy the CreatorSprout team

Launch coverage based on Edge0’s November 12, 2026 announcement, its published model cards and independent reporting on the release. We have not tested Edge0-145B hands-on. Every benchmark and speed figure below is Edge0’s own claim from its internal testing, not a result we measured. No affiliate links in this article.

A 145-billion-parameter model running on a laptop sounds like a typo, and that is exactly why this release got attention: the Hacker News thread drew hundreds of points within a day, and the GitHub repo passed 10,000 stars. The trick is a mixture-of-experts design. Only about 10 billion of the 145 billion parameters fire on any given token, so the model behaves more like a 10B model at inference time while keeping the knowledge of a much bigger one. Edge0 is not the first to try this, but it is the first to ship it open at this scale with a framework built for personal hardware. The honest question is not whether the model exists — the weights are public — it is whether the laptop story holds up once independent testers run it.

What was announced

Edge0 announced Edge0-145B on Thursday, November 12, 2026, together with a smaller sibling, Edge0-120B, plus its own unified inference framework and deployment tools. The company calls Edge0-145B the first open-source on-device model at the 145-billion-parameter scale. The weights are on Hugging Face, the framework code is on GitHub under an Apache 2.0 licence, and the release follows Edge0-35B, a phone-sized model the company shipped on September 10, 2026 that it says ran at 37+ tokens a second on an iPhone 18 Pro Max using about 2.5GB of memory. Edge0-145B is the same idea scaled roughly four times: frontier-scale weights, consumer hardware, no account, no usage limits, and no data leaving your machine.

The architecture: 145B total, 10B active

Mixture of experts, 48 layers. Edge0-145B is a latent mixture-of-experts model with 512 experts per layer, of which 10 routed experts plus 1 shared expert are active on each token — about 10.12 billion active parameters out of 145 billion total. That is why a model this large can even be contemplated on a laptop: most of the weights sit idle on any given step.
Native 4-bit training. Edge0 says the model was trained in its on-device architecture from the outset, rather than compressed down from a cloud model. One wrinkle: the Hugging Face metadata lists Alibaba’s Qwen3.8-Flash-Next as the base model, and Edge0’s own release-scope file says the actual training base revision is unverified, so the “trained from scratch for on-device” story is the company’s claim, not a settled fact.
256K context window. Edge0 lists context from 128K to 256K depending on configuration, which puts long documents and large codebases in reach on local hardware — if the memory budget allows.
Speed claims. Edge0 reports prefill above 1,100 tokens a second, decode above 80 tokens a second, and typical power of about 120W on a compatible laptop. All of these are the company’s own measurements; no independent run has been published yet.

The 29GB headline vs the 61.5 GiB fine print

This is the part of the launch that matters most, and it is where the coverage splits. Edge0’s blog headline promises 29GB peak memory on a laptop such as a high-memory MacBook Pro. But Edge0’s own model card tells a fuller story: for the NVFP4 build it measures 23.25 GiB of GPU memory plus 38.27 GiB of host RAM, about 61.5 GiB in total, and the weights themselves are 47.79 GiB for NVFP4 and 40.81 GiB for the Q3_K build. The 29GB figure is the GPU-side peak only. The shipped engine also documents CUDA, not Mac, which sits awkwardly next to the MacBook Pro example in the announcement. The practical read: plan for a machine with 48GB of unified memory or a workstation with 64GB, not a 32GB laptop. That is still a remarkable number for a 145B model, but it is not the number in the headline.

Benchmarks: strong, but all Edge0’s own

DeepSWE 1.1: 64.6. Edge0 reports 64.6 for Edge0-145B, against 58.7 for Qwen3.8-Flash-Next and 54.4 for DeepSeek-V4-Flash-0731, on its own internal evaluations.
Terminal-Bench 2.1: 86.52. Edge0 reports 86.52, ahead of DeepSeek-V4-Flash at 82.7 and Qwen3.8-Flash-Next. Terminal-Bench tests real agentic coding work in a terminal, so this is the number coding users will care about most.
Six-benchmark mean: 72.8. Across DeepSWE 1.1, Terminal-Bench 2.1, HLE, GPQA-Diamond, IFBench and AA-LCR, Edge0 claims a mean of 72.8, against 71.7 for Qwen3.8-Flash-Next.
Read them as the vendor’s case. Every one of these figures comes from Edge0’s internal evaluations, and the comparisons with Qwen, DeepSeek and Claude are internal too. Edge0 says a technical report with the full setup and reproduction instructions is being released. Until independent testers run the model on real laptops, these scores are a promise, not a verdict.

Cost and licence: free, with commercial limits

Free to download, free to run. There is no per-token API fee because the model runs on your own machine, and it works offline without an account. The cost is the hardware and the electricity, which Edge0 itself notes.
The weights are not fully open. Edge0-145B is released under Edge0’s own community licence (edge0-community-model-1.0, dated November 12), not a standard open licence. Three uses need a separate written licence: selling, renting or licensing the weights; operating a model service for third parties financed by fees, subscriptions, advertising or sponsorship where customers control inference inputs; and offering an independent AI coding or office-productivity product financed the same way. Personal use, research and most internal business use are not in that list, but read the licence before building a product on it.
The framework is Apache 2.0. The inference framework and deployment tools on GitHub carry the standard Apache 2.0 licence, which is the genuinely open part of the release.

Where you can run it

Weights: Hugging Face under Edge0, including the Edge0-145B-q3k build.
Framework and tools: GitHub at Edge0-AI/edge0, Apache 2.0.
Platforms: Edge0 says day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA RTX Spark, RTX 4090 and RTX 5090 hardware. Note the caveat above: the shipped engine documents CUDA, so Windows and Linux on NVIDIA hardware are the best-supported path today, and Mac support is worth verifying before you commit to it.
Hardware target: laptops with 48GB of memory, or workstations with 64GB, per Edge0’s own target-device guidance.

What it means for creators and developers

Private coding help, no API bill. If the Terminal-Bench numbers hold up, a local 145B model that codes at 80+ tokens a second is a genuine alternative to paying per token for cloud coding models — and your code never leaves the machine. That is the use case to test first.
Offline work that keeps working. No account, no usage limits, no internet required once the weights are downloaded. For creators on the move, in the field, or handling sensitive client material, that is a real workflow change.
Try it as a weekend experiment, not a purchase decision. The download is free, so the cost of testing is a few dozen gigabytes of disk and an evening. Run your hardest real task through it — a real codebase, a real long document — and compare the output against the cloud model you use today before changing anything.
Check the licence before you ship anything. If you plan to wrap the model in a paid product or a paid service, the community licence’s written-licence requirement applies. Personal projects and internal use are the safe ground.
For comparison, see how the cloud flagships stack up. Our coverage of Claude Fable 5.2, Gemini 4 Carbon and Kimi K3.1 Pro shows what the paid API route costs and scores — Edge0-145B is the bet that local hardware can now match much of it for free.

The honest caveats

We haven’t tested it. Everything here comes from Edge0’s announcement, its published model cards and independent reporting on the release. The speed, memory and benchmark figures are Edge0’s, not results we measured, and no independent laptop run has been published yet.
The memory headline is misleading. The 29GB figure is GPU memory only. Edge0’s own card measures about 61.5 GiB total for the NVFP4 build, so a 48GB machine is the realistic floor. Do not buy hardware for this model based on the headline alone.
The benchmarks are the vendor’s own. DeepSWE, Terminal-Bench and the six-benchmark mean all come from Edge0’s internal testing, with no independent verification and no technical report yet. Wait for the second wave of measured runs before rebuilding a workflow around it.
The “trained from scratch” story is contested. The Hugging Face metadata names Qwen3.8-Flash-Next as the base model, and Edge0’s own release-scope file says the training base revision is unverified. That does not make the model bad, but it does mean the origin story deserves a question mark.
The licence is not open source in the usual sense. The community licence reserves paid model services and paid coding or office products for a separate written licence. If your plan involves charging for access to the model, you need Edge0’s permission, not just the download.

Verdict

Edge0-145B is the most interesting open-model release of the month, and the most honestly caveated one. A free 145B-parameter model with 10B active per token, a 256K context window and a framework built for personal hardware is a real step toward frontier AI that lives on your desk instead of in someone else’s data centre. But the 29GB headline is only the GPU half of a roughly 61.5 GiB total, the benchmarks are all Edge0’s own, and the licence is not the open licence the word “open-source” usually implies. If you have a 48GB laptop or a strong NVIDIA workstation, download it and test it on your hardest real task this weekend. If you don’t, wait for the independent runs — they will tell us whether this is a laptop model or a workstation model with a laptop headline.

FAQ

What is Edge0-145B?

Edge0-145B is an open-weights language model released by Edge0 on November 12, 2026. It has 145 billion total parameters, of which about 10 billion are active on each token, and it is designed to run on a personal computer instead of in a data centre. The weights are free to download from Hugging Face under Edge0’s own community licence.

How much does Edge0-145B cost?

The weights and the inference framework are free to download, there is no per-token API fee, because the model runs on your own machine. You still pay for the hardware and electricity, and Edge0’s community licence requires a separate written licence for some commercial uses, including selling or renting the weights or running a paid model service.

Can Edge0-145B really run on a laptop?

Edge0 says yes on a well-equipped laptop, quoting 29GB peak memory and 80+ tokens a second. But its own model card measures about 61.5 GiB of total memory for the NVFP4 build (23.25 GiB of GPU memory plus 38.27 GiB of host RAM), so a 48GB laptop is the realistic target. No independent laptop run has been published yet.

How good is Edge0-145B at coding?

Edge0’s own internal benchmarks put it at 64.6 on DeepSWE 1.1 and 86.52 on Terminal-Bench 2.1, ahead of Qwen3.8-Flash-Next and DeepSeek-V4-Flash on those tests. Those are the vendor’s numbers from its own testing, not independently verified results, and a technical report with the full setup is still to come.

Where can I download Edge0-145B?

The model weights are on Hugging Face under Edge0 (Edge0-145B-q3k and related builds), and the inference framework and deployment tools are on GitHub at Edge0-AI/edge0 under an Apache 2.0 licence. Edge0 says day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA RTX Spark, RTX 4090 and RTX 5090 hardware.