Edge0, a company that builds on-device AI models, open-sourced Edge0-145B on November 12. Its blog post calls the model the first open-source on-device model with 145 billion parameters and says it runs at 80 tokens per second or more on a laptop. The 145B model’s peak memory is listed at 29GB.
Below: what Edge0-145B claims, what hardware reports say it needs, how the speeds compare, what is unverified, and a checklist for anyone who wants to try it. The checklist is The Jolt Journal’s own. The Jolt Journal did not run the model.
Edge0-145B at a glance
Every figure below is from Edge0’s announcement unless noted. The company says the numbers come largely from internal testing.
| Item | Edge0’s claim |
|---|---|
| Release date | November 12, 2026 |
| Size | 145 billion parameters, about 10 billion active per step |
| Design | Sparse mixture of experts; 64-expert groups share two large matrices |
| Peak memory | 29GB (a later paragraph says 28GB) |
| Speed | 1,100+ tokens per second prefill; 80+ tokens per second decode |
| Power | About 120W typical |
| Context | 128K to 256K tokens |
| Sibling model | Edge0-120B |
- Edge0 says it trained the model for on-device use from the start, rather than compressing it afterward.
- The weights are on Hugging Face under the Edge0 organization, and the inference framework is on GitHub.
What hardware does Edge0-145B need?
This is where accounts differ. The 29GB figure is a peak memory number. It is not a download size and not a minimum for a machine.
| Topic | What is reported | Source |
|---|---|---|
| Memory tiers | Hot experts in GPU memory; about 40GB expert library in system memory; about 48GB of vocabulary data read from SSD | Edge0 |
| Download size | Q3_K file 40.81 GiB; NVFP4 file 47.79 GiB | ByteIota |
| GPU | NVIDIA with CUDA 13; test rig used a 24GB RTX 5090 D | Zeniteq |
| Operating systems | Linux and Windows listed; macOS not listed for the 145B model | Zeniteq, ByteIota |
| Edge0’s own list | macOS, Windows and Linux on day one, including Apple M-series | Edge0 |
The accounts do not agree on Macs. Edge0’s post names Apple M-series chips. ByteIota and Zeniteq said the 145B and 120B models do not run on macOS yet. ExplainX said the models are effectively unsupported on Mac for now. The Jolt Journal’s read of the GitHub README showed Apple Silicon steps only for the smaller 35B and 8B models, and it may be out of date. Mac owners should check the repository before downloading.
The site’s report on the M6 MacBook Pro covers Apple’s newest laptop line.
How fast is Edge0-145B?
Edge0’s post gives laptop speeds without naming the laptop. Zeniteq published a table of Edge0’s own desktop-GPU results.
| Hardware and format | 16K prefill | Chat decode | Agent decode |
|---|---|---|---|
| RTX 5090 D 24GB, Q3_K | 3,554.9 tokens/s | 99.7 tokens/s | 121.2 tokens/s |
| RTX 5090 D 24GB, NVFP4 | 3,165.7 tokens/s | 79.6 tokens/s | 98.3 tokens/s |
| RTX 4090 24GB, Q3_K | 3,179.4 tokens/s | 69.9 tokens/s | 91.7 tokens/s |
Zeniteq said the laptop claims of 1,100+ and 80+ tokens per second are not tied to a configuration and should not be compared with this table. Speed measures also leave out model loading, tool calls and reasoning tokens.
Edge0-145B benchmark scores
Edge0 published six scores. They are the company’s own evaluations, and its post says some rows mix sources.
| Benchmark | Edge0-145B | Edge0-120B |
|---|---|---|
| DeepSWE 1.1 | 64.6 | 53.6 |
| Terminal-Bench 2.1 | 86.52 | 81.3 |
| HLE | 35.4 | 31.0 |
| GPQA-Diamond | 90.4 | 87.4 |
| IFBench | 80.6 | 79.3 |
| AA-LCR | 79.3 | 81.0 |
| Six-test mean | 72.8 | 68.9 |
Edge0 put the mean for Qwen3.8-Flash-Next at 71.7 and for DeepSeek-V4-Flash-0731 at 70.1. On Terminal-Bench 2.1 it listed Qwen3.8-Flash-Next at 83.3 and Claude-Opus-4.6 at 78.2. ExplainX said it found no independent tests cited. Commenters on Hacker News, per Daily.dev, asked whether the scores came from the 29GB build or from full-precision weights.
Where Edge0-145B came from and its license
- Base model. Zeniteq said both models are derived from Qwen3.8-Flash-Next. Edge0’s post describes native on-device training. The two accounts differ. The site’s report on Qwen3.8-Flash-Next covers that model.
- License. The GitHub repository lists Apache-2.0, per The Jolt Journal’s read of it. Edge0’s post names no license for the weights. ExplainX said the weights ship with their own model cards and advised checking upstream terms.
- “Open source.” A Hacker News commenter, per Daily.dev, said the release is open weights rather than fully open source.
For another open-weights release, see the site’s report on Mistral Large 4.
Checklist before you run AI locally on a laptop
These are The Jolt Journal’s suggestions, not Edge0’s guidance.
- Check the GPU. Reports point to an NVIDIA card with CUDA 13. A laptop without one may not qualify.
- Budget three kinds of memory. Plan for GPU memory, system RAM and SSD space separately. The 29GB figure covers only one of them.
- Check disk space. The Q3_K file alone is 40.81 GiB, per ByteIota.
- Read the license. Check the model card before any commercial use.
- Test your own task. Vendor benchmarks do not predict results on your files. Try a small job first.
- Watch cooling and power. Edge0 reports about 120W. Laptop thermals are untested by third parties.
- Know the cloud alternative. Local runs avoid usage fees, but hardware and electricity still cost money. The site’s guide to a Gemini API spend cap covers the cloud side.
What is not known about Edge0-145B
- Which laptop produced the 80+ tokens per second figure.
- Whether Edge0-145B runs on Apple Silicon at all.
- Whether the benchmark scores apply to the quantized builds people can download.
- How speed holds up on slow storage, since the model streams experts from SSD.
- The active-parameter count of Edge0-120B.
- Whether the 29GB or 28GB peak is correct.
Editor’s take
Editor’s take: Edge0-145B is a large open model that Edge0 says can run on one high-end machine, and the memory-streaming design is the interesting part. The “laptop” label, though, depends on hardware that most laptops lack. Until independent testers reproduce the speeds and scores, the safest reading is that a 24GB NVIDIA GPU, plenty of RAM and fast storage are the realistic starting point.
This article is for general information, based on Edge0, GitHub, Zeniteq, ByteIota, ExplainX and Daily.dev as of November 13, 2026. The Jolt Journal did not run the model. Specifications and availability may change.
