Edge0-145B could be installed on launch day — its biggest advantage over a research-only model. But "installable" does not mean "installable on any computer": there is a gap between the headline "runs on 29GB" and the shipping documentation's requirements, and Mac support is a clear discrepancy. This article works through the requirements, the choice between the two builds, the install steps and the Mac situation, following the GitHub v0.2.0 release notes and the Hugging Face model cards. Note: this site has not run these steps on real hardware — commands and paths follow the GitHub README.
| Item | Requirement (GitHub v0.2.0 release notes) |
|---|---|
| GPU | NVIDIA GPU, compute capability 7.5+, 16 GiB VRAM minimum |
| System RAM | 20 GiB minimum (measured whole-machine demand is higher — see the table below) |
| OS | Linux or Windows (build instructions target NVIDIA GPUs with CUDA 13); x64 and arm64 packages available |
| CUDA | CUDA Toolkit 13 |
| Disk | Weight files of 40.81 or 47.79 GiB, plus an expert library of about 40GB (system memory) and vocabulary data of about 48GB (SSD) |
16GB of VRAM is the floor, not the recommendation: the model card's measured residency (RTX 5090 D 24GB) is NVFP4 at 23.25 GiB VRAM + 38.27 GiB host RAM, and Q3_K at 23.14 GiB VRAM + 31.28 GiB host RAM. A 16GB card gets through the door by the official bar, but the measured environment uses a 24GB card; 20 GiB of system RAM is the floor, while measured use is 31–38 GiB. Check memory before installing — too little VRAM fails outright, and too little system RAM fails while loading the expert library.
| Item (RTX 5090 D 24GB, vendor-measured) | Q3_K | NVFP4 |
|---|---|---|
| Weight file | 40.81 GiB | 47.79 GiB |
| VRAM used | 23.14 GiB | 23.25 GiB |
| Host RAM used | 31.28 GiB | 38.27 GiB |
| Whole machine, total | About 54.4 GiB | About 61.5 GiB |
| Agent decode | 121.2 tok/s | 98.3 tok/s |
| 16K prefill | 3,554.9 tok/s | 3,165.7 tok/s |
| Repository | Edge0/Edge0-145B-q3k | Edge0/Edge0-145B-NVFP4 |
Default to Q3_K. Smaller file (40.81 vs 47.79 GiB), less host RAM (31.28 vs 38.27 GiB), faster (121.2 vs 98.3 tok/s decode) — it wins on all three. NVFP4's advantage is precision: the 4-bit floating-point format should in theory preserve more detail than 3-bit quantization, but no official score difference between the two builds has been published, so the "precision for speed" trade cannot be quantified. Start with Q3_K unless you have a specific reason; try NVFP4 side by side for precision-sensitive work.
Step 1 — check the environment. Match the table above: GPU (compute capability 7.5+, 16 GiB+ VRAM), system RAM (20 GiB+, 32 GiB+ recommended), CUDA Toolkit 13, Linux or Windows. Keep at least 50 GiB of disk free for the weight file, plus about 48GB for vocabulary data.
Step 2 — download the weights. Download the chosen build from Hugging Face: Q3_K (Edge0/Edge0-145B-q3k) or NVFP4 (Edge0/Edge0-145B-NVFP4). The weights are not bundled with the edge0-infer installer — the v0.2.0 release notes say so explicitly; the 145B weights must be downloaded separately.
Step 3 — install edge0-infer. Get edge0-infer v0.2.0 (release notes dated Nov 11, 2026) from GitHub Edge0-AI/Edge0; Windows x64 and arm64 packages are available. The framework and deployment code are Apache 2.0. Follow the repository README's build instructions — the build target is Linux or Windows with an NVIDIA GPU and CUDA 13.
Step 4 — start the local server or chat. edge0-infer offers a local server (serve) and a chat (chat) entry point; the local server speaks the OpenAI- and Anthropic-compatible Messages APIs — so once it is running, existing OpenAI / Anthropic client code can talk to it, and tool calling goes through the same local server. Vision input needs the strata-vision helper (not built from the Edge0 tree), so check its status during setup.
Users who would rather not use the command line can take the Edge0 App desktop application: per the official blog, its core capabilities work offline with no account and no usage quota. But the v0.2.0 Edge0 Chat desktop packages are Windows-only. So "download the app and open it" currently works for Windows users only; Linux users take the edge0-infer command line, and Mac users — see the next section.
This is the most disputed point of the launch, and it needs stating plainly: the official blog claims day-one support for macOS / Apple M-series / RTX Spark / RTX 4090 / 5090, and markets it with a MacBook Pro as the example; but the GitHub README's platform table (read Nov 13 by multiple sources) lists no macOS support for the 145B/120B models — the build instructions target NVIDIA GPUs with CUDA 13 on Linux or Windows, and the v0.2.0 Edge0 Chat desktop packages are Windows-only. The 57.0 tok/s figure on Apple M2 Ultra (Q3_K, agent decode) is a Metal speed measurement, not proof of macOS availability. Conclusion: do not buy hardware or make plans on "a MacBook Pro can run it" at launch; wait until the README adds macOS support or a Mac Edge0 App ships. The Chinese critical piece newtoken.link put it directly — "Edge0's 145-billion-parameter model cannot run on Mac at launch" — and as of Nov 20, 2026, that judgement still holds.
Reference: GitHub — Edge0-AI/Edge0 · Hugging Face — Edge0/Edge0-145B-q3k model card · Hugging Face — Edge0/Edge0-145B-NVFP4 model card · Edge0 — Official blog "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer" (rewritten and organized by ModelBrief; facts per official sources)