Edge0-145B Puts a 145-Billion-Parameter Model on One GPU — Open-Weight, Not Open-Source
Edge0 released Edge0-145B and Edge0-120B on 12 November 2026, with its inference framework and deployment tools, and the headline claim is a striking one: a 145-billion-parameter model that runs on a single GPU. The vendor calls Edge0-145B "the world's first open-source on-device AI model with 145 billion parameters". That "open-source" label is the vendor's wording, and it is not accurate for the weights. The weights are open-weight, under a community licence with real restrictions, and they are not open-source. That distinction, the memory numbers behind the single-GPU claim, and the fact that every benchmark is the vendor's own internal evaluation, are the three things to keep in mind.
What Edge0-145B is
Edge0-145B has 145 billion total parameters, with about 10.12 billion active per token. The model is a latent sparse mixture-of-experts model, in which only a small part of the model is active for any given token. The architecture is 48 layers, with 512 experts per layer, and 10 routed experts plus 1 shared expert active per token. The context window is 256,000 tokens. The model was trained natively at 4 bits. Two builds are offered: Q3_K, at 40.81 GiB, and NVFP4, at 47.79 GiB.
The base model is derived from Qwen3.8-Flash-Next, from Alibaba. The Hugging Face model card acknowledges Qwen3.8-Flash-Next and Strata, while the announcement blog does not name Qwen. The model is a derivative of Qwen3.8-Flash-Next, and the engine is a fork of Strata, and it should not be presented as a from-scratch model. That is a matter of accuracy, not a criticism: building on an existing base model is a normal way to build a model, and the derivative relationship is stated openly on the model card. Edge0-120B was released alongside it, on the same day, with the same framework and tools.
![]() |
High-end desktop PC with a large graphics card visible through a glass side panel glowing with subtle lighting, on a desk in a dark room (AI-generated illustration, credit: INORYVA)
The single-GPU claim, and what the memory numbers really mean
The headline figures come from Edge0's own measurements, and they should be read as such. On an RTX 5090 D with 24 GB of memory, using the Q3_K build, Edge0 reports 121.2 tokens per second for agent decode, 3,554.9 tokens per second for 16K prefill, and 86.5 on Terminal-Bench 2.1. The same build was also measured on an RTX 4090 with 24 GB, at 91.7 tokens per second for agent decode, and on RTX Spark N1X and Apple M2 Ultra hardware using Metal. The blog also claims 29 GB peak memory and more than 80 tokens per second for decode. None of those figures has been independently verified.
The memory figure needs the most care, because it is the easiest to misread. The blog's "29 GB peak memory" is GPU-side only. It should not be presented as total system memory. The Hugging Face model card's own measured residency tells a fuller story: for the NVFP4 build, 23.25 GiB of video memory plus 38.27 GiB of system RAM, or about 61.5 GiB in total; for the Q3_K build, 23.14 GiB of video memory plus 31.28 GiB of system RAM. A laptop needs far more than 29 GB in total. The 29 GB figure is a GPU-side peak, not the whole requirement.
| Build | Video memory (measured) | System RAM (measured) | Total residency |
|---|---|---|---|
| Q3_K (40.81 GiB) | 23.14 GiB | 31.28 GiB | About 54.4 GiB, plus storage |
| NVFP4 (47.79 GiB) | 23.25 GiB | 38.27 GiB | About 61.5 GiB, plus storage |
![]() |
The benchmarks — all the vendor's own
The benchmarks for Edge0-145B are all Edge0's own internal evaluations, with no independent verification. The blog itself states: "Edge0, Qwen, DeepSeek, and Claude results come from Edge0's internal evaluations", and it adds that "the complete evaluation setup, hardware environment, and reproduction instructions are being released with our technical report" — in other words, not yet released. Until that technical report is available, the numbers cannot be checked by anyone outside Edge0, and they should be reported as Edge0's numbers, not as established facts about the model.
With that label attached, the numbers Edge0 reports are: DeepSWE 1.1, 64.6; Terminal-Bench 2.1, 86.52; HLE, 35.4; GPQA-Diamond, 90.4; IFBench, 80.6; AA-LCR, 79.3; and a mean across the six benchmarks of 72.8. The comparisons in the blog table put those numbers alongside Qwen3.8-Flash-Next at 58.7 on DeepSWE and 83.3 on Terminal-Bench, DeepSeek-V4-Flash-0731 at 54.4 and 82.7 on the same two, and Claude-Opus-4.6 in its Max configuration at 91.3 on GPQA and 78.2 on Terminal-Bench. The comparison models' memory footprints in the blog table are estimates, not measured figures.
Open-weight, not open-source
The licence is the part of this release that most needs to be stated plainly, because the vendor's own language points the other way. The weights are not open-source. They are open-weight, under the Edge0 Community Model License 1.0, and Qwen materials remain subject to the Qwen Community License 1.0. Both sets of terms must be followed. Open-source, in its ordinary meaning, implies a set of freedoms without further permission being needed. Open-weight means the weights are available, under a licence that may, and in this case does, restrict some of those freedoms. The weights are available. The freedoms are limited.
The licence permits personal use, research, modification, noncommercial sharing, internal use and commercial use of generated outputs. But selling or commercially bundling the model, operating a commercial model service, or offering an independent commercial coding or office assistant using it all require a separate written licence from Edge0. Those are the core commercial uses that an open-source licence would normally allow without further permission, and they are the uses this licence withholds. The framework code is different: edge0-infer is under the Apache-2.0 licence, and Strata retains its MIT licence. Those permissive licences apply to the code. They do not apply to the weights, and the code being open-source does not make the weights open-source.
![]() |
Laptop on a desk running a local AI chat interface on screen, a coffee cup beside it, warm home office light (AI-generated illustration, credit: INORYVA)
Availability, platforms, and what is left out
The weights are available on Hugging Face, at the Edge0-145B-q3k page and in the NVFP4 build, and the inference framework and deployment tools are on GitHub, at the Edge0-AI Edge0 repository. The Edge0 App, a desktop application, is available through GitHub Releases. The announcement says day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA RTX Spark, 4090 and 5090 hardware. But the Hugging Face card states that the Apple M2 Ultra measurements use Metal and "do not establish macOS release availability", and secondary checks found no Mac runtime or guidance. This article does not promise Mac users that they can run the model today.
The blog's cost and duration demo claims are vendor anecdotes, not verified facts, and are omitted here. For context on the wider open-model landscape, see our coverage of the Impeccable GitHub skill, Mistral Large 4, Kimi K3.1 Pro and the flagship AI model comparison. Edge0-145B sits in a different place: a large model meant to run on one GPU, under a community licence, with its benchmarks still to be independently checked.
Key Takeaways
- Edge0 released Edge0-145B and Edge0-120B on 12 November 2026, with its inference framework and deployment tools.
- Edge0-145B has 145B total parameters and about 10.12B active per token, with a 256K context, trained natively at 4 bits; the model is a derivative of Qwen3.8-Flash-Next, not a from-scratch model.
- The single-GPU figures are Edge0's own measurements: on an RTX 5090 D 24 GB, Q3_K, 121.2 tokens per second agent decode and 3,554.9 tokens per second 16K prefill.
- "29 GB peak memory" is GPU-side only; measured total residency is about 61.5 GiB for NVFP4 and about 54.4 GiB for Q3_K, plus storage.
- All benchmarks are Edge0's own internal evaluations, not independently verified; the technical report with the evaluation setup has not yet been released.
- The weights are open-weight under the Edge0 Community Model License 1.0, not open-source; selling or commercially bundling the model, operating a commercial model service, or offering an independent commercial coding or office assistant using it requires a separate written Edge0 licence.
- Do not promise Mac users they can run it today: the blog claims day-one macOS support, but the model card says the M2 Ultra measurements do not establish macOS release availability.
Conclusion
Edge0-145B is a genuine step towards large models running on a single GPU, and it is also a release that needs to be read with its labels attached. The memory the model actually needs is much larger than the headline peak figure suggests, the benchmarks are the vendor's own and have not yet been independently checked, and the weights are open-weight under a community licence with real commercial restrictions — not open-source. Read with those three corrections in mind, the release is what it is: a 145-billion-parameter model that can run on one GPU, in a system with substantial RAM and storage, under a licence that lets you use it for yourself and for research, and that asks you to come back to Edge0 before you sell it or sell access to it.



Comments
Post a Comment