DEV Community

Cover image for Edge0-145B Open Weights: RTX GPU vs Mac, RTX Wins
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

Edge0-145B Open Weights: RTX GPU vs Mac, RTX Wins

For builders choosing hardware, the Edge0-145B open weights are an RTX story, not a Mac story. On November 12, 2026, Edge0 released the 145-billion-parameter model with weights, an inference engine and a desktop app. The company reports 29GB peak memory, but the GitHub requirements set a floor of an NVIDIA GPU with at least 16 GiB of VRAM plus 20 GiB of system RAM. A 12 GB card, a Mac or a free Colab T4 will not run it, so the verdict is a single RTX card with 16 GiB or more, on Linux or Windows.

TL;DR

  • Edge0-145B and Edge0-120B were released on November 12, 2026, alongside the edge0-infer v0.2.0 engine and the Edge0 Chat app.
  • Hardware floor: NVIDIA GPU, CUDA 13, compute capability 7.5 or higher, 16 GiB VRAM, 20 GiB system RAM.
  • macOS is not supported for the 145B and 120B models in the v0.2.0 release, despite the launch post suggesting otherwise.
  • Benchmark scores are Edge0's own and are strongest on coding and terminal tasks.
  • The weights use a community licence, not Apache 2.0, and Qwen's upstream licence still applies.

Which should you pick: an RTX GPU or a Mac?

Pick the RTX GPU. The platform table in the GitHub README lists Linux and Windows with NVIDIA hardware for the 145B and 120B models. Metal and macOS support applies to the smaller Edge0-35B tier. The launch post's FAQ says macOS works on day one, but independent write-ups from enjoy.dev and explainx.ai noted that this is not reflected in the release itself.

If you already own a Mac, the practical option is Edge0-35B, which Edge0 says runs under 2.5GB on a phone. It is a different model with a different capability level, so treat it as an alternative and not a substitute.

What does Edge0-145B need to run locally?

The requirements, as published in the repository, are an RTX 20, 30, 40 or 50 series card (compute capability 7.5 or higher), CUDA 13, 16 GiB of GPU memory and 20 GiB of system RAM.

The "29GB" figure refers to peak memory in Edge0's own measurement, which counts the GPU and RAM tiers together. Third-party testing by eesel.ai on a 24GB card found 23.14 GiB of GPU memory plus 31.28 GiB of system RAM for the Q3_K build, about 54.4 GiB in total. These figures are not contradictory. They describe different configurations, so budget for more RAM than the headline suggests if you want headroom.

Cards that miss the floor:

  • RTX 3060 12GB: under the 16 GiB VRAM minimum.
  • Free Colab T4: compute capability 7.5, but its 15 GiB of VRAM is about 1 GiB short.
  • Free Kaggle P100: compute capability 6.0, below the 7.5 requirement.

How does Edge0-145B fit on one GPU?

The model is a latent mixture-of-experts derived from Alibaba's Qwen3.8-Flash-Next. According to the launch post, it has 145B total parameters with about 10B active per token, 512 experts per layer across 48 layers, ten experts selected per token and one shared expert. Edge0 says the two builds, Q3_K and NVFP4, were trained natively at 4 bits.

The memory trick is a three-tier layout. Frequently used experts stay in GPU memory, a library of about 40GB sits in system RAM, and about 48GB of vocabulary data is read from SSD on demand. Groups of 64 experts share two large matrices, and each expert keeps only a small delta. The GPU computes cached experts while the CPU works in parallel, and speculative decoding spreads weight reads across candidate tokens.

The practical consequence is that fast system RAM and a fast SSD matter alongside the GPU. A 16 GiB card with slow storage will not behave like the figures below.

How fast is Edge0-145B on real hardware?

All of these numbers come from Edge0's README testing on an RTX 5090 D, so treat them as vendor measurements.

Setup (Q3_K build) 4K prefill Chat decode Agent decode
RTX 5090 D, full 24GB 3,236.7 tok/s 130.2 tok/s 144.6 tok/s
RTX 4090 24GB 2,637.5 tok/s 95.0 tok/s 104.9 tok/s
RTX 5090 D limited to 16 GiB VRAM and 20 GiB RAM 430 tok/s 62.4 tok/s 73.3 tok/s

The constrained row is the one to read if you own a 16 GiB card. It is slower than the full-memory runs but still comfortably above reading speed. The NVFP4 build is slower under the same limit, at 51.3 and 59.9 tokens per second for chat and agent decode, per the GitHub README.

Do the benchmark scores hold up?

They are internal evaluations, and no independent reproduction has been published as far as the sources I reviewed show. Edge0 reports a six-benchmark mean of 72.8 for the 145B model against 71.7 for Qwen3.8-Flash-Next and 70.1 for DeepSeek-V4-Flash-0731. On individual tests it reports 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond (close to Claude-Opus-4.6 (Max) at 91.3), and 80.6 on IFBench against 62.5 for the same Opus model, per the launch post. It does concede ground on Humanity's Last Exam, scoring 35.4 against 40.0 for Claude-Opus-4.6 (Max).

The pattern is credible: a model tuned for agentic coding and terminal work, with less margin on hard knowledge questions. Run your own tasks before committing a workflow to it. A note on scope, using our own measurement: first-party DataForSEO pricing run completed 2026-11-12, keyword corpus n=695, found that only 72 out of 695 AI-tooling keywords clear a winnable search bar, which is why verdict pages like this one are rare rather than filler. For comparison with other self-hostable releases, see our notes on LongCat 2.5 and Kimi K3.1 Pro.

Is Edge0-145B really open source?

The artifacts are open, but the licence is not Apache. The model weights fall under the Edge0 Community Model License 1.0, dated November 12, 2026, which permits personal use, research, modification, noncommercial sharing, internal use and commercial use of generated outputs. Selling or bundling the model, running a commercial model service, or offering an independent commercial coding or office assistant needs a separate written licence from Edge0. The terms are in the Hugging Face repository.

Upstream rights remain. The NOTICE file states that Edge0 does not claim the whole model belongs to it, and the Qwen Community License 1.0 continues to apply, as writeble.com also points out. The engine is Apache-2.0 and Strata is MIT, as explainx.ai summarises. If you plan a product, read both licences first. Our Runway Praxis 1 GPU verdict covers a similar licence-and-hardware check.

How do you set it up?

  1. Use a Linux or Windows machine with an RTX 20, 30, 40 or 50 series GPU of 16 GiB or more and at least 20 GiB of RAM. Install CUDA Toolkit 13.
  2. Clone the Edge0 repository and build edge0-infer, or take the Windows package from release v0.2.0.
  3. Download a checkpoint from Hugging Face: Edge0/Edge0-145B-q3k or Edge0/Edge0-145B-NVFP4. The 120B variants are smaller but have the same GPU floor.
  4. Run edge0-infer serve for a local OpenAI-compatible and Anthropic Messages-compatible endpoint, then point your coding agent at localhost. Alternatively use edge0-infer chat or the Edge0 Chat desktop app.

If offline use is your main goal, our piece on OMNII-1 running on a laptop shows a lighter-weight path. Builders interested in post-training a Qwen base can also look at Arete 27B.

FAQ

Q: Can Edge0-145B run on a MacBook?
A: Not in the current release. The platform table and the v0.2.0 notes list Linux and Windows with NVIDIA GPUs for the 145B and 120B models. Only Edge0-35B has Metal support on macOS.

Q: What is the minimum GPU for Edge0-145B open weights?
A: An NVIDIA card with CUDA 13, compute capability 7.5 or higher and at least 16 GiB of VRAM, plus 20 GiB of system RAM, according to the project README.

Q: Can I use Edge0-145B commercially?
A: Partly. Commercial use of generated outputs is allowed, but selling or bundling the model, running a commercial model service, or shipping a commercial assistant needs a written licence from Edge0. The Qwen Community License also still applies.

Q: Does the free Colab tier work?
A: No. The free T4 has about 15 GiB of VRAM, which is below the 16 GiB floor. Kaggle's P100 fails the compute capability requirement.

Q: Should I choose Edge0-145B or Edge0-120B?
A: Choose 145B if your hardware meets the floor, since Edge0 reports higher scores on most benchmarks. The 120B has the same GPU floor but a smaller footprint, so it mainly helps if RAM or disk is tight.

Top comments (0)