ERA OF MINDS

Edge0-145B requirements: can you run it on a laptop or a Mac? (what the release files say)

Short answer

Edge0-145B runs on your own machine only if you have Linux or Windows, an NVIDIA GPU with at least 16 GiB of memory and CUDA 13, at least 20 GiB of system RAM, and about 98 GiB of free disk for the model folder. There is no macOS build for the 145B model yet, even though the launch blog mentions day-one macOS support; on a Mac, the Edge0 app runs only the 8B and 35B models. The weights are free on Hugging Face. If your GPU sits right at 16 GiB, the smaller Edge0-120B runs about 1.7 times faster on the same budget.

Era of Minds cover: Edge0-145B on a laptop? A retyped copy of Samuel Zeng's launch post (298K views, 884 bookmarks): Cloud-level intelligence on your laptop. Below, what the release files say today: Linux or Windows with an NVIDIA GPU and CUDA 13, at least 16 GiB GPU memory and 20 GiB system RAM, about 98 GiB download for Q3_K, no macOS build for 145B yet, and the Mac app runs only the 8B and 35B models.
Figure: Samuel Zeng's launch post retyped by us (not a screenshot), next to what the GitHub README, the v0.2.0 release notes and the Hugging Face model cards say, checked November 13, 2026.

Samuel Zeng (@SamuelZengML on X) announced Edge0-145B and Edge0-120B at 18:00 UTC on November 12 (10 PM in Dubai): "Cloud-level intelligence on your laptop. Free, offline, and fast enough for real vibe coding." By 10:30 Dubai time the next morning the post had 297,796 views, 975 likes and 884 bookmarks (X API). X counted 13 posts about Edge0 on November 11, 429 on November 12 and 297 more in the first six and a half hours of November 13 (UTC). The Edge0-AI/Edge0 GitHub repo had 10,533 stars.

Most of those replies ask the same thing: will it run on my laptop? The launch post and the release files give slightly different answers, so we read the README, the release notes, both model cards and the docs and put the hard numbers in one place.

What is Edge0-145B?

A free, open-weights model built to run on one consumer GPU instead of a cloud API. From the model page in the repo:

  • Size: 145 billion parameters in total, but only about 10.12 billion work on each token. It is a mixture-of-experts model: 48 layers, 512 experts per layer, 10 picked per token plus one shared.
  • Lineage: derived from Qwen3.8-Flash-Next and trained natively at 4 bits.
  • Context: 256K tokens; the server defaults to 131,072.
  • Builds: Q3_K (40.81 GiB of weights) and NVFP4 (47.79 GiB), released as separate Hugging Face folders.
  • Sibling: Edge0-120B, 8.64 billion active per token, smaller and faster, scores lower.
  • Job: agentic coding. The server speaks OpenAI- and Anthropic-compatible endpoints on your own machine.

The trick that lets a 145B model fit is streaming. The hot experts sit in GPU memory, an expert library of about 40GB sits in system RAM, and a roughly 48GB vocabulary table stays on the SSD and is read as needed, per the launch blog.

What are the Edge0-145B requirements?

WhatRequirement (release files, Nov 13)
Operating systemLinux or Windows (x64 or Arm64). Not macOS for 145B/120B
GPUNVIDIA, CUDA 13. Windows x64 build: compute capability 7.5 or newer (RTX 20 / 30 / 40 / 50). Windows Arm64 build: compute capability 12.1 only (RTX Spark N1X)
GPU memoryAt least 16 GiB
System memoryAt least 20 GiB
DiskAbout 98 GiB for the Edge0-145B-q3k folder (the PLE table alone is 47.68 GiB), about 105 GiB for NVFP4, about 86 GiB for Edge0-120B-q3k
SoftwareNVIDIA CUDA Toolkit 13. To build from source: CMake 3.24+, Ninja, a C++17 compiler, Python 3.10+ and uv

Sources: the v0.2.0 release notes, the edge0-145b run guide and file sizes from the Hugging Face API. One detail matters for planning. The model card says the 16 GiB VRAM plus 20 GiB RAM setup is a "test limit, not a minimum system requirement." It was an RTX 5090 D 24 GB capped by software. Without that cap, the same card held the Q3_K build as 23.14 GiB of VRAM plus 31.28 GiB of RAM. So 16 and 20 will work, but more is faster.

Can Edge0-145B run on a Mac?

Not with the files released so far. This is where the launch materials disagree with themselves:

  • The launch blog says: "day-one support covers macOS, Windows, and Linux, including Apple M-series chips," and gives a MacBook Pro as an example laptop.
  • The Edge0-AI/Edge0 GitHub README platform table puts a dash next to macOS for edge0-145b and edge0-120b. Only the 35B model runs on macOS (Metal and CPU).
  • Both model cards include an Apple M2 Ultra row in the speed table (60.9 tokens per second in chat), then add that the M2 Ultra measurements "do not establish macOS release availability."
  • The Edge0 macOS guide has you build the Edge0 app and then pick Edge0 8B or Edge0 35B. It needs macOS 14 and an M3 or later chip.

So a Mac owner can try Edge0 today, just not the 145B model. On a MacBook Air M3, Edge0 measured about 1.7 GB of resident memory for 8B and about 4 GB for 35B (device requirements page). If Mac support for 145B ships, the README platform table is the place it will show up first.

Will Edge0-145B run on my laptop?

Run three checks before you download almost 100 GiB:

  1. GPU memory. On Windows, open Task Manager, Performance, GPU and read "Dedicated GPU memory." On Linux, run nvidia-smi. You need 16 GiB or more on an NVIDIA card. Integrated graphics and AMD cards are not supported by the released engine.
  2. System RAM. 20 GiB is the tested floor. With 32 GB you are close to what the uncapped test used (31.28 GiB).
  3. Free SSD space. About 98 GiB for Q3_K. A fast SSD helps, because the server reads the embedding table from disk by default (ple-io=direct).

The launch blog's headline number, "29GB peak memory," is Edge0's internal measurement, and the blog also says it enables use "on a 48GB laptop." Read that as a gaming or workstation laptop with a big NVIDIA GPU, not a thin ultrabook.

How fast is Edge0-145B?

Machine (Edge0's tests)BuildChat decodeAgent decode16K prefill
RTX 5090 D 24 GBQ3_K99.7121.23,554.9
RTX 4090 24 GBQ3_K69.991.73,179.4
RTX Spark N1X 60 GiBNVFP440.146.0922.4
Apple M2 Ultra (Metal, not released)Q3_K60.957.0800.3
Same 5090 D capped to 16 GiB VRAM + 20 GiB RAMQ3_K36.844.9335

Tokens per second, from the Edge0-145B-q3k model card. The last row is the one most laptop owners should look at: around 37 tokens per second in chat is still faster than you read, but long prompts take a while at 335 tokens per second of prefill.

How do I install and run Edge0-145B?

Windows, the easy way. Install the NVIDIA CUDA Toolkit 13. Download edge0-chat-0.2.0-windows-x64.zip from the v0.2.0 release notes (Arm64 users take the arm64 zip). Download a model folder from Hugging Face, for example Edge0/Edge0-145B-q3k, and keep it as one folder. Open Edge0 Chat and choose that folder. Model weights are not inside the app zip.

Linux (or Windows) from the command line:

  1. Clone the repo: git clone https://github.com/Edge0-AI/Edge0.git edge0
  2. Build the engine, edge0-strata, with CUDA 13 by following "Build (Linux, NVIDIA)" in edge0-strata/EDGE0.md. It creates build/strata and a Python environment in .venv.
  3. Build edge0-infer with the cpu preset (it needs no GPU backend of its own for these models), per edge0-infer/docs/getting-started.md.
  4. Download the model folder, for example with the Hugging Face CLI: hf download Edge0/Edge0-145B-q3k --local-dir Edge0-145B-q3k. Pass the folder for one model, not a parent folder holding several.
  5. Start the server from edge0-infer/: build/cpu/bin/edge0-infer serve --model /path/to/Edge0-145B-q3k. Loading can take a while; the default wait is up to 1,800 seconds. You are ready when it prints Listening: http://127.0.0.1:8000 model=edge0_145b.
  6. Test it: curl localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{"model":"edge0_145b","messages":[{"role":"user","content":"hi"}],"max_tokens":64}'

For an interactive session use edge0-infer chat --model /path/to/Edge0-145B-q3k instead, but don't run chat and serve on the same GPU at once; the docs warn against it. Because the server answers OpenAI- and Anthropic-style requests on 127.0.0.1:8000, a coding tool that lets you set a custom base URL can point at it. We have not tested a specific agent with it yet.

Does Edge0-145B really beat Claude Opus 4.6?

On one coding benchmark, by Edge0's own run. Here is the full comparison from the model page in the repo:

BenchmarkEdge0-145BEdge0-120BQwen3.8-Flash-NextClaude-Opus-4.6 (Max)
Terminal-Bench 2.186.575.383.378.2
DeepSWE 1.164.651.358.7no score
GPQA-Diamond90.487.491.791.3
HLE35.431.035.940.0
IFBench80.679.381.362.5

The launch blog says the Edge0, Qwen, DeepSeek and Claude results all "come from Edge0's internal evaluations," and that the full setup and reproduction steps will come with a technical report. Note also that Edge0 beats its own base model, Qwen3.8-Flash-Next, by just 3.2 points on Terminal-Bench. Until outside testers publish numbers, treat "beats Opus" as a claim on one benchmark, not a general result.

Edge0-145B or Edge0-120B: which should I download?

Edge0-145B (Q3_K)Edge0-120B (Q3_K)
Active parameters per token10.12B8.64B
Weights40.81 GiB27.98 GiB
Uncapped on RTX 5090 D 24 GB23.14 GiB VRAM + 31.28 GiB RAM22.93 GiB VRAM + 18.40 GiB RAM
Chat / agent speed at 16 GiB VRAM + 20 GiB RAM36.8 / 44.9 tok/s62.4 / 73.3 tok/s
Terminal-Bench 2.1 (Edge0)86.575.3

Our rule of thumb from these numbers: on a 16 GB GPU with 32 GB of RAM or less, start with 120B. With a 24 GB card and 64 GB of RAM, go straight to 145B.

Is Edge0-145B free, and can I use it at work?

The download is free and the framework code is Apache-2.0. The weights use the Edge0 Community Model License 1.0, which allows personal use, research, education, modification, noncommercial sharing, internal use and commercial use of the outputs. Selling or bundling the model, or running a commercial model service on it, needs a separate license (info@edge0.ai). Qwen's Community License still applies on top, since the model is derived from Qwen3.8-Flash-Next. Running it locally also costs electricity; the blog puts typical power at about 120W.

Our take

Edge0-145B is a real release you can download today, with unusually honest model cards that state their own limits. The catch is in the wording. "On your laptop" means a laptop with a 16 GB-plus NVIDIA GPU, plenty of RAM and 100 GiB of free SSD, and "beats Opus" is one benchmark from Edge0's own run. If you already own that hardware, it is a free way to stop paying per token for routine coding work. If you are on a Mac, wait for the README platform table to change.

If your reason for looking is a cloud bill, also read how we cut tokens with Headroom for Claude Code, and what happened when MiniMax ended its free API but left the weights out in our MiniMax M3.1 Flash post. And if your limit is a Claude plan, our explainer on the Claude free full reset shows how to get a full refill before you switch tools.

FAQ: Edge0-145B requirements, Mac support and setup

What is Edge0-145B?

Edge0-145B is a free, open-weights AI model from Edge0, released November 12, 2026 with a smaller sibling, Edge0-120B. It is a mixture-of-experts model with 145 billion parameters, of which about 10.12 billion work on each token, a 256K context window and builds in Q3_K and NVFP4. It is derived from Qwen3.8-Flash-Next and tuned for agentic coding, and it is meant to run on one consumer NVIDIA GPU instead of a cloud API.

What are the Edge0-145B requirements?

Per the GitHub release notes and docs (checked November 13, 2026): Linux or Windows, an NVIDIA GPU with CUDA 13 and at least 16 GiB of GPU memory (compute capability 7.5 or newer, so RTX 20-series and up on Windows x64), at least 20 GiB of system memory, and about 98 GiB of disk for the Q3_K model folder (about 105 GiB for NVFP4). 16 GiB plus 20 GiB is the tested floor, not the comfortable setup.

Can Edge0-145B run on a Mac?

Not with the released files today. The launch blog says day-one support covers macOS and Apple M-series chips, but the GitHub README lists no macOS support for edge0-145b and edge0-120b, and the model card says its Apple M2 Ultra speed test does not establish macOS release availability. On a Mac, the Edge0 app runs the smaller Edge0 8B and 35B models on M3 or later.

Can Edge0-145B run on a laptop?

Yes, if the laptop has an NVIDIA GPU with at least 16 GiB of video memory, 20 GiB or more of system RAM, CUDA 13 and room for a roughly 98 GiB download. A thin laptop with integrated graphics or an 8 GB GPU will not run it. The 29GB peak memory figure in the launch blog is Edge0's internal measurement, not a promise for every machine.

How fast is Edge0-145B?

Edge0's own table: 99.7 tokens per second in chat and 121.2 in agent work on an RTX 5090 D 24 GB (Q3_K), 69.9 and 91.7 on an RTX 4090, and 36.8 and 44.9 when the same card is limited to 16 GiB of VRAM and 20 GiB of RAM. These are the company's numbers; independent tests were not out yet on November 13.

Does Edge0-145B really beat Claude Opus 4.6?

On one benchmark, by Edge0's own test. It reports 86.5 on Terminal-Bench 2.1 against 78.2 for Claude-Opus-4.6 (Max), and the blog says the Claude numbers also come from Edge0's internal evaluations. Opus scores higher on GPQA-Diamond (91.3 vs 90.4) and HLE (40.0 vs 35.4). Treat it as a claim until outside benchmarks land.

Is Edge0-145B free for commercial use?

The weights are free to download and the framework code is Apache-2.0. The Edge0 Community Model License 1.0 allows personal use, research, modification, internal use and commercial use of the outputs, but not selling or bundling the model or running a commercial model service without a separate deal (info@edge0.ai). Qwen's license terms also apply because the model is derived from Qwen3.8-Flash-Next.

Should I pick Edge0-145B or Edge0-120B?

Pick 120B if your machine is close to the minimum. Its Q3_K weights are 27.98 GiB instead of 40.81 GiB, and on the limited 16 GiB VRAM plus 20 GiB RAM setup it runs at 62.4 tokens per second in chat versus 36.8 for 145B. Pick 145B for quality: Edge0 reports 86.5 vs 75.3 on Terminal-Bench 2.1 and 64.6 vs 51.3 on DeepSWE 1.1.

Sources

FREE WEEKLY AI BRIEF

Free. Unsubscribe anytime. Free gift: 50 AI prompts that actually work, sent the moment you join. More about Decoded