Edge0-145B: a 145-billion-parameter open model that runs on a laptop at 80 tokens a second
Edge0 has open-sourced Edge0-145B, a 145-billion-parameter model it says runs on a 48GB laptop at more than 80 tokens a second — with benchmark scores that, on the company’s own tests, rival frontier cloud models. Every headline number is the vendor’s own, and none has been independently verified yet.
- Edge0 open-sourced Edge0-145B and Edge0-120B on November 12, 2026, with its inference framework, deployment tools and Edge0 App; the company calls the 145B model the first open-source on-device model at that scale.
- Reported laptop performance: 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, about 120W typical power, and a 128K–256K context window — figures from Edge0’s own testing.
- Of the 145B parameters, only about 10B are active on each inference step; the design shares large matrices across groups of 64 experts instead of giving every expert a full set.
- Benchmarks from Edge0’s internal evaluations, not independently verified: 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond and 80.6 on IFBench — a 72.8 mean across six benchmarks, against 71.7 for Qwen3.8-Flash-Next.
- The weights are on Hugging Face under Edge0’s own community licence; the framework code is Apache 2.0 on GitHub; day-one support covers macOS, Windows and Linux.
- The full footprint is bigger than the 29GB headline: a roughly 40GB expert library sits in system memory and about 48GB of vocabulary data lives on the SSD.
On Wednesday, a small team called Edge0 made one of the bolder claims of the year in local AI: a 145-billion-parameter model that runs on a laptop. Edge0-145B, open-sourced on November 12 alongside a smaller 120B sibling, is billed as the first open-source on-device model at that scale — and the company says it brings coding, reasoning and agent capabilities that rival frontier cloud models to a personal computer, at more than 80 tokens a second.
If the numbers hold up, it would move the frontier for what a single machine can do without the cloud. The catch is the sourcing: every headline figure comes from Edge0’s own testing, and the technical report that would let anyone reproduce the benchmarks is not out yet. The weights and code are public, so verification can start now.
What Edge0 says the model can do#
The headline numbers are about coding and reasoning. On DeepSWE 1.1, a benchmark of real software-engineering work, Edge0-145B scores 64.6, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4. On Terminal-Bench 2.1 it scores 86.52, again ahead of both. On GPQA-Diamond, a hard science test, it scores 90.4 — close to DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 (Max) at 91.3. On IFBench, instruction following, it scores 80.6.
| Model | DeepSWE 1.1 | Terminal-Bench 2.1 | GPQA-Diamond | IFBench | HLE | AA-LCR |
|---|---|---|---|---|---|---|
| Edge0-145B | 64.6 | 86.52 | 90.4 | 80.6 | 35.4 | 79.3 |
| Qwen3.8-Flash-Next | 58.7 | 83.3 | 91.7 | 81.3 | 35.9 | 79.1 |
| DeepSeek-V4-Flash-0731 | 54.4 | 82.7 | 90.8 | 79.2 | 33.8 | 79.7 |
| Claude-Opus-4.6 (Max) | — | 78.2 | 91.3 | 62.5 | 40.0 | 78.0 |
Across the six benchmarks, Edge0-145B averages 72.8 against 71.7 for Qwen3.8-Flash-Next, the closest on-device comparison. Every one of these scores comes from Edge0’s internal evaluations — the company says so itself, and the full evaluation setup, hardware details and reproduction instructions are promised with a technical report that is not yet out. Until independent runs exist, the table is a claim about what the model can do, not a settled measurement.
How a 145B model fits on a laptop#
The trick is not one idea but four, designed together. The model is a mixture-of-experts design in which large matrices are shared across groups of 64 experts, so each expert keeps only a small set of parameters of its own. Of the 145 billion parameters in total, only about 10 billion take part in computation on any single step.
The other three layers are about moving data. Training was built for local execution from the start, rather than compressing a cloud model after the fact. Loading pipelines GPU memory, system memory and SSD access so reading and computation overlap: frequently used experts stay in GPU memory, a roughly 40GB expert library sits in system memory, and about 48GB of vocabulary data lives on the SSD, read on demand. Scheduling runs the CPU and GPU in parallel and uses speculative decoding to verify several candidate tokens at once.
The honest version of the memory story is more qualified than the headline. The 29GB peak-memory figure applies when running through Edge0’s own inference framework, and the machine still needs room for the expert library and the vocabulary data on disk. Edge0 names laptops with 48GB of memory — a MacBook Pro, for example — and 64GB workstations as the target hardware. That is still a remarkable claim: before September 2026, Edge0 says, memory limits kept most on-device models below 10 billion parameters. Apple’s on-device model is about 3B parameters; Google’s Gemini Nano came in 1.8B and 3.25B versions.
What you can actually download#
The release is a full stack, not just weights. The model weights are on Hugging Face, the inference framework and deployment tools are on GitHub under an Apache 2.0 licence, and there is an Edge0 App for running it. Day-one support covers macOS, Windows and Linux, including Apple M-series chips and Nvidia’s RTX Spark, RTX 4090 and RTX 5090. The company says the app’s core capabilities work offline, without an account and without a usage quota.
One licensing detail matters: the code is Apache 2.0, but the weights carry Edge0’s own community licence, not a standard open licence — anyone planning to ship a product on the model should read it first. Edge0 has done this before at smaller scale: its Edge0-35B phone model reportedly reaches 37+ tokens a second on an iPhone 18 Pro Max using about 2.5GB of memory.
Why it matters — and what to verify#
The reason a 145B local model matters is agents. Edge0’s demonstrations show the model planning a task, writing code, running it, debugging failures and checking the result visually — building a 3D scene of Manhattan in Blender, producing an explainer video about a flying Tesla Roadster, and creating playable games. Those workflows take dozens of rounds of revision, and running them locally means no per-token cloud bill for every retry, no data leaving the machine, and no dependence on a connection.
That is the promise; the verification is still ahead of us. The benchmark table is vendor-run, the technical report is pending, and the demos are the company’s own. Coverage noted the GitHub repository passing 10,000 stars within a day — the idea has landed with developers, but stars are not benchmarks. The number that will settle this is one nobody outside Edge0 has produced yet: an independent run on a real 48GB laptop.
What to watch#
Three things will tell us whether the claim survives contact with reality. First, the technical report: Edge0 promises the full evaluation setup, hardware environment and reproduction instructions — the document that turns a vendor table into something checkable. Second, independent benchmarks: whether anyone reproduces 80+ tokens a second and the 29GB peak on a real machine. And third, whether other labs follow: if a 145B model can really live on a laptop, the on-device ceiling that held for years has just moved.
Sources#
- Edge0 — “Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer” (November 12, 2026)
- Glonce — “Edge0 open-sources Edge0-145B, a 145B model for laptops” (November 13, 2026)
- Tokenstrom — “Edge0-145B: An open 145-billion-parameter model that’s meant to run on your laptop” (November 13, 2026)
- Spectrum Web — “Edge0 releases 145B open model for a single laptop” (November 13, 2026)
- Medium (Aakash Das) — “Edge0 Says Its 145B Model Runs on a Laptop. The Weights Are Free” (November 13, 2026)