# Edge0-145B open-sources a 145B model that runs on a laptop

> Edge0-145B is an open-source 145B model that Edge0 says runs on a 48GB laptop with 29GB peak memory and 80+ tokens/s. Here is what founders should...

- URL: https://tools.launchllama.co/blog/edge0-145b-open-source-145b-model-runs-on-laptop
- Published: 2026-11-13
- Publisher: Launch Llama (https://tools.launchllama.co)

TL;DR
- Edge0 open-sourced Edge0-145B and Edge0-120B along with its inference framework and deployment tools.
- Reported laptop numbers: 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, about 120W typical power, 128K to 256K context.
- Benchmark claims come from Edge0's own internal evaluations. The full setup and reproduction steps are coming with a technical report.

## What happened

Edge0 released the weights for Edge0-145B and Edge0-120B, together with its unified inference framework and deployment tools. It calls Edge0-145B the first open-source on-device model with 145 billion parameters. Only about 10B of those parameters take part in each inference step.

On the hardware side, Edge0 reports 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode and roughly 120W typical power on a compatible laptop such as a MacBook Pro. It says a 48GB laptop is enough. The page gives 29GB in one table and 28GB in another comparison for the same laptop run, so treat the exact peak as roughly 28 to 29GB.

On Edge0's internal evaluations, Edge0-145B scores 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond and 80.6 on IFBench. Claude-Opus-4.6 (Max) is ahead on HLE (40.0 vs 35.4) and GPQA-Diamond (91.3 vs 90.4). Edge0 says the full evaluation setup and reproduction instructions are not out yet.

## What this means for founders

- **Cloud bills:** Edge0 says local inference has no cloud token or API fees, and that data can stay on the machine and work offline. It also says hardware and electricity still cost money, so this is a trade, not free compute.
- **Hardware gate:** The stated target is a 48GB laptop. If your team is on 16GB or 32GB machines, the source does not show a way to run the 145B model on them.
- **Agent-heavy workflows:** Edge0 demos multi-round work such as Blender scenes and games that run dozens of plan, run and check loops. Long loops are where per-token cloud pricing adds up, so that is the use case to test first.
- **Data-sensitive products:** If customer code or documents cannot leave the device, an open-weight model with a full inference stack is worth a look. You can evaluate or fine-tune it and integrate it into your own product.
- **Risk:** The comparisons are Edge0's own, memory figures for rival models are estimates not yet measured, and the license terms are not in the text I had. Check all three before you build on it.

## How to try it today

- Check your machine first. Edge0 says a 48GB laptop is enough, and its reported run used a compatible laptop such as a MacBook Pro.
- From the Edge0 blog post, open the links for the model weights, the inference framework and the Edge0 App. The App is the quickest way to try it without setting up the stack.
- Read the license on the weights before you ship anything. The source text I had does not state it.
- Test it on one real task from your own work, not a demo. Time it and compare quality against the cloud model you use now.
- Watch for the technical report with the evaluation setup and reproduction steps before you trust the benchmark numbers.Example test prompt (my own, not from Edge0):
Read this repo, find the failing test, fix it, run the tests, and tell me what you changed.

## Related tools on Launch Llama

Tools from the Launch Llama directory that fit with this news:

ToolWhat it doesWhy it matters hereWhat's my local ai?For AI engineers: What's my local ai? detects your hardware and recommends local models so you run LLMs privately without cloud costs.Detects your hardware and recommends local models, a quick way to see whether your machine fits a model this size.ModelBrewFor developers: ModelBrew fine-tunes open-source LLMs on your data with continual learning so you own a custom model without vendor lock-in.Edge0 says you can fine-tune the released models; ModelBrew is a paid option for fine-tuning open-source LLMs on your own data.Router by RampFor engineers: Router by Ramp routes every LLM request to the lowest-cost model that meets your needs so you cut inference costs and simplify API management.If the local setup does not fit your hardware, this routes cloud requests to the lowest-cost model that meets your needs.LLM ReferenceFor engineers: LLM Reference compares 1,700+ models across 130+ providers so you pick the right LLM and lowest price in seconds.Compares models and prices across providers, useful for checking what you would pay in the cloud against running locally.

[Browse all tools](https://tools.launchllama.co/tools) or [list your own tool](https://tools.launchllama.co/submit).

## Llama's take

I would try this now if you already have a 48GB laptop and pay real money for agent loops on a cloud model. A local model with no per-token fees is worth a weekend test, especially for coding tasks where you can check the output yourself.

I would wait if you are on smaller hardware, or if you need independent benchmarks before switching. Every score above comes from Edge0's internal runs, and the reproduction details are still pending. The memory comparisons against rival models are also estimates.

The thing to watch is the technical report. If outside teams reproduce the 80+ tokens/s and the coding scores on ordinary laptops, this changes the local-versus-cloud math for small teams.

## FAQ

### How much memory does Edge0-145B need?

Edge0 reports 29GB peak memory on a compatible laptop, and says this enables local use on a 48GB laptop. Another table on the same page lists 28GB, so expect roughly 28 to 29GB.

### Is Edge0-145B open source?

Edge0 says it is open-sourcing Edge0-145B and Edge0-120B, along with its inference framework and deployment tools. The license terms were not in the text I reviewed, so check them before commercial use.

### Is Edge0-145B better than Claude?

It is mixed, and the numbers are Edge0's own. It beats Claude-Opus-4.6 (Max) on Terminal-Bench 2.1 (86.52 vs 78.2) and IFBench (80.6 vs 62.5). Claude is ahead on HLE (40.0 vs 35.4) and GPQA-Diamond (91.3 vs 90.4).

## Sources

- [Edge0-145B: Frontier AI Capabilities on Your Own Computer](https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer) (edge0.ai)
- [Hacker News discussion](https://news.ycombinator.com/item?id=50475713)

---

More for agents: https://tools.launchllama.co/llms.txt · https://tools.launchllama.co/catalog.md · https://tools.launchllama.co/mcp
