Edge0 Says a 145B Model Runs on a 29GB Laptop, but Its Repo Lists 40.8 GiB of Weights and 48GB of Embeddings on Disk
Edge0's 145B open-weight model does run locally, but the 29GB laptop claim hides 40.8 GiB of weights, a 24GB GPU and 48GB of embeddings on disk.

Edge0 released a 145-billion-parameter open-weight model on 12 November and said it runs on a “compatible personal laptop” with 29GB of peak memory and more than 80 tokens per second. The company, which also shipped a 120B sibling, calls Edge0-145B the first open-source on-device model at that scale.
The catch is in Edge0’s own repository. The Q3_K build is 40.81 GiB of weights, the model card lists 23.14 GiB of video memory plus 31.28 GiB of system memory when experts are held in residency, and the blog says about 48GB of vocabulary data sits on the SSD and is read on demand.
That does not make the release fake. It runs, and Edge0 publishes speed tests from an RTX 5090, an RTX 4090, an RTX Spark machine and an Apple M2 Ultra. It does mean the 29GB number in the FAQ is not the memory a buyer needs, and “laptop” covers a small set of expensive ones.
Quick answer: Edge0-145B is a real Apache 2.0 release with 145B total parameters and 10.12B active per token. It needs an NVIDIA GPU with CUDA 13 on Linux or Windows, or Apple Silicon on macOS. The Q3_K weights are 40.81 GiB, residency needs about 54 GiB of VRAM and RAM combined, and the embedding table adds roughly 48GB on disk. The advertised 29GB is a measured peak in one configuration, not a minimum requirement.
What Edge0 actually shipped
Edge0 open-sourced two latent mixture-of-experts models, Edge0-145B and Edge0-120B, together with its inference framework and deployment tools. Both are derived from Qwen3.8-Flash-Next, and both are licensed Apache 2.0.
That lineage matters. This blog has covered Qwen’s own releases, including Qwen 3.8-27B under Apache 2.0, and Edge0 is best understood as a compression and serving layer on top of that family rather than a new pretraining run.
The 145B model has 48 layers, 512 routed experts per layer plus one shared expert, and a top-10 routing rule. Only 10.12B parameters take part in each token, which is why the active compute is closer to a 10B model than a 145B one.
Edge0 also claims a 256K context, a 1,100 tokens per second prefill rate, about 120W typical power and a 1.26B multi-token-prediction head for speculative decoding. Those are company figures, not independent measurements.
The 29GB claim and the 40.8 GiB build
Edge0’s blog says the model runs “on a compatible personal laptop, such as a MacBook Pro, without a server” at 29GB peak memory. Its FAQ goes further and says the hardware requirement is “a compatible laptop with at least 29GB of memory.”
The repository tells a more specific story. The Q3_K build is 40.81 GiB, and the NVFP4 build is 47.79 GiB. With residency on, the Q3_K weights occupy 23.14 GiB of VRAM and 31.28 GiB of host RAM, which is 54.42 GiB of combined memory for weights alone.
Then there is the vocabulary table. It has 51.2B parameters in fp8, and the blog says about 48GB of it lives on the SSD. Add that to the weight build and the download is close to 90GB before the first token. That total is my arithmetic from Edge0’s two published figures, not a number the company states.
The 29GB figure is best read as a measured peak in a streaming configuration, where most experts stay on disk and host memory. The repository even warns that its 16 GiB VRAM plus 20 GiB RAM test is “a measured test budget, not a general minimum requirement.” The FAQ dropped that caveat.
What the numbers look like side by side
| Model | Total / active parameters | Weights | Memory to run | Mean score, six tests | Agent decode |
|---|---|---|---|---|---|
| Edge0-145B Q3_K | 145B / 10.12B | 40.81 GiB | 23.14 GiB VRAM + 31.28 GiB RAM | 72.8 | 121.2 tok/s on RTX 5090 |
| Edge0-120B Q3_K | 120B / 8.64B | 27.98 GiB | 22.93 GiB VRAM + 18.40 GiB RAM | 68.9 | 144.6 tok/s on RTX 5090 |
| Qwen3.8-Flash-Next | Not disclosed here | About 70GB at 4-bit, estimated | About 70GB | 71.7 | Not listed |
| Qwen3.8-27B | 27B dense | About 18GB at 4-bit, estimated | About 18GB | 66.1 | Not listed |
| DeepSeek-V4-Flash | Not disclosed here | About 87GB at 4-bit, estimated | About 87GB | 70.1 | Not listed |
The comparison is not apples to apples. Edge0’s memory figures are measured peaks, while the three rivals are 4-bit estimates that Edge0 says “remain to be measured.” A 145B model with 10B active parameters should be compared with other sparse models on quality and with dense models only on cost.
The benchmark table is vendor-run
Edge0 reports 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond, 80.6 on IFBench, 79.3 on AA-LCR and 35.4 on HLE. It says those results beat Qwen3.8-Flash-Next on coding and agent work, and trail it on GPQA, HLE and IFBench.
The blog discloses that “Edge0, Qwen, DeepSeek, and Claude results come from Edge0’s internal evaluations.” No third party has reproduced the numbers, and the technical report with the evaluation setup was not linked in the post at the time of writing.
Against Claude Opus 4.6, the gap is mixed. Edge0 wins IFBench 80.6 to 62.5 and AA-LCR 79.3 to 78.0, loses HLE 35.4 to 40.0, and is close on GPQA. Claude has no published DeepSWE 1.1 score, so the headline coding comparison is incomplete. Treat the table as a vendor claim with a plausible shape, not a verdict.
What you actually need to run it
The supported platforms are Linux or Windows with an NVIDIA GPU and CUDA 13, or macOS on Apple Silicon with Metal. On macOS only the Q3_K build works, and the context is capped at 32,768 tokens, not 256K.
For the 80 to 121 tokens per second range, the hardware is a desktop RTX 5090 with 24GB, an RTX 4090 with 24GB, an RTX Spark N1X with 60GiB, or an M2 Ultra. The M2 Ultra manages 60.9 tokens per second in chat and 57.0 in agent mode, well below the 5090.
The trade-off is familiar. Ornith-1.5-122B, also a sparse model with about 10B active parameters, needs a single 80GB GPU at 4-bit. Edge0’s claim is that its engine cuts that requirement, and the engine is the product.
A “laptop” in this context means a machine with a 24GB mobile RTX 5090 or a MacBook Pro with 64GB or more of unified memory and a fast SSD. Those cost several times an entry-level developer laptop, and the 48GB embedding file makes storage a real constraint.
Confirmed vs not confirmed
| Claim | Status | Source |
|---|---|---|
| Edge0-145B and 120B were released with open weights | Confirmed | Edge0 blog and GitHub release notes |
| The weights are Apache 2.0 | Confirmed | Repository license file |
| 145B total and 10.12B active parameters | Confirmed | Model card |
| Q3_K build is 40.81 GiB | Confirmed | Model card |
| It runs on a laptop with at least 29GB of memory | Not confirmed | FAQ says so, model card gives 54.42 GiB resident |
| 80+ tokens per second on a laptop | Partly confirmed | 121.2 tok/s on RTX 5090; 57.0 on M2 Ultra |
| World’s first open-source 145B on-device model | Not confirmed | Company claim, no independent check |
| 256K context on macOS | Not confirmed | Model card caps macOS at 32K |
| Beats Claude Opus 4.6 overall | Not confirmed | Vendor-run results, mixed per test |
What it means for Indian developers
The appeal is real for teams that cannot send code or customer data to a US cloud. A model that runs offline with no per-token bill changes the economics of long agent loops, code review and document work, especially for regulated clients in health, finance and government.
The cost moves to hardware. A 24GB GPU laptop or a 64GB MacBook Pro is a capital expense, and downloading roughly 90GB of weights and embeddings is a bandwidth event in a country where many offices still run on shared connections. Electricity at 120W is trivial next to that.
The technique is not new. Colibri ran a 744B model on 25GB of RAM by streaming layers, and Tencent’s Hy3 put a 295B model on one GPU. Edge0’s contribution is packaging the trick with trained weights, a supported engine and a permissively licensed release.
There is also a licensing advantage. Apache 2.0 allows commercial use and fine-tuning, and Edge0 says the toolchain covers macOS, Windows and Linux. Indian startups that already build on Qwen can test Edge0 as a drop-in local option without a new licence review.
For most teams, the practical comparison is not Edge0 against a frontier API. It is Edge0 against a 27B model that fits in 18GB and runs on hardware they already own. The 145B model wins on agent and coding benchmarks, but only if the machine is there.
FAQ
Can Edge0-145B run on a 29GB laptop?
Not on 29GB of total memory. The Q3_K weights are 40.81 GiB, and Edge0’s own residency figures need 23.14 GiB of VRAM plus 31.28 GiB of RAM. The 29GB number is a measured peak in a streaming setup, and the model card calls its constrained test a budget, not a minimum.
Is Edge0-145B really open source?
The weights and inference code are published under Apache 2.0, and the models are derived from Qwen3.8-Flash-Next. That permits commercial use and modification, subject to the licence terms and any upstream notices.
How fast is Edge0-145B in practice?
Edge0 reports 121.2 tokens per second in agent mode and 99.7 in chat on an RTX 5090, 91.7 and 69.9 on an RTX 4090, and 57.0 and 60.9 on an M2 Ultra. A constrained 16 GiB VRAM plus 20 GiB RAM test reaches 44.9 and 36.8.
Does it run on a MacBook?
Yes, on Apple Silicon with Metal and only with the Q3_K build. The context is capped at 32,768 tokens on macOS, and the first start writes an extra draft-model file into the model folder, which must be writable.
Is it better than Qwen3.8-Flash-Next?
On Edge0’s own tests it is ahead on DeepSWE 1.1 and Terminal-Bench 2.1, roughly level on AA-LCR, and behind on GPQA-Diamond, HLE and IFBench. No independent evaluation has confirmed those results.
Sources
- Edge0, “Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer”, 12 November 2026. edge0.ai
- Edge0, model card for Edge0 145B and 120B, repository documentation, accessed 14 November 2026. github.com
- Edge0, release notes, 13 November 2026. github.com
- Edge0, Apache License 2.0. github.com