Edge0 announced Edge0-145B and Edge0-120B on November 12, 2026, alongside inference software and deployment tools for running large AI models locally. The company pitches the release for coding, reasoning and agent workloads that would otherwise depend on cloud services. Its announcement claims laptop operation with approximately 29GB of peak memory and generation above 80 tokens per second.
The release’s most useful qualification appears in the GitHub documentation. Release notes dated November 13 describe both models as latent mixture-of-experts models derived from Qwen3.8-Flash-Next. Their currently documented deployment path uses a single NVIDIA GPU on Linux or Windows with CUDA 13.
That distinction matters: downloadable weights, a laptop benchmark and a supported installation path are separate things. Edge0 has published artifacts developers can evaluate, but its broad laptop positioning should not be read as confirmation that every advertised platform supports these two models today.
What Is Actually Available to Download
Edge0’s Hugging Face organization page lists four checkpoints across the two model sizes:
| Model | Published checkpoint repositories | Intended use |
|---|---|---|
| Edge0-145B | Edge0/Edge0-145B-q3k and Edge0/Edge0-145B-NVFP4 | Local coding and agent workloads |
| Edge0-120B | Edge0/Edge0-120B-q3k and Edge0/Edge0-120B-NVFP4 | Similar workloads with smaller weight files |
Edge0 says both models were trained natively at four bits. The released inference formats are Q3_K and NVFP4, giving developers two quantized representations to evaluate.
The opened model listings do not provide exact downloadable byte totals, so there is no supported basis here for assigning a specific gigabyte size to any checkpoint. Check the repository’s actual files and selected revision before planning storage or downloading.
Both models use 48 layers, 512 experts and a routing setting of K=10, according to the repository’s model table. They are derived models, not releases trained entirely from scratch independently of Qwen3.8-Flash-Next.
I’d treat the availability of both weights and an inference stack as the practical news. Developers can inspect the implementation and evaluate their own workloads instead of relying entirely on a hosted demo. Commercial adoption still calls for checking the applicable model and software license terms.
Why “145B” Does Not Tell You the Download Size
Edge0 describes its largest model as having 145 billion total parameters, with approximately 10 billion participating in computation during each inference step. Those figures answer different questions.
- Total parameters describe the model’s stated architectural scale.
- Active parameters describe how much of that architecture participates in a particular inference step.
- Stored weight bytes depend on the representation, quantization, shared matrices and supporting data.
- Runtime memory includes resident weights, caches, working buffers and other allocations.
A sparse model can activate only selected experts for a token without keeping every expert’s full computation active. Sparsity alone, however, does not make inactive weights disappear from storage.
Edge0 describes another reduction mechanism in its architecture explanation: groups of 64 experts share two large matrices, while each expert retains a smaller set of parameters describing its differences. The company also pipelines access across GPU memory, system memory and SSD storage.
That combination is why multiplying the name by a bit width can mislead. As an illustrative calculation, 145 billion independently stored values at exactly four bits each would occupy 72.5GB in decimal units, before overhead. That is not an estimate of Edge0’s download size: the calculation ignores its shared-matrix representation and actual file format.
Likewise, a reported 29GB memory peak cannot be substituted for a download size or a VRAM requirement. A streaming system can keep some weights outside GPU memory, while long-context inference introduces additional cache requirements.
My view is that teams should budget storage, system RAM and VRAM separately. “145B runs locally” is an architectural claim; a useful deployment specification needs all three resource figures.
The Supported Hardware Path Is More Specific Than “Laptop”
The latest platform support table separates the new models from Edge0’s previous releases:
| Component | Edge0-145B and Edge0-120B |
|---|---|
| Operating systems | Linux and Windows |
| Accelerator | NVIDIA GPU |
| CUDA requirement | CUDA 13 |
| Model engine | edge0-strata, a fork of Strata |
| Shared interface | edge0-infer serve and edge0-infer chat |
| Documented macOS, iOS or Android support | Not listed for these models |
The shared edge0-infer interface selects an engine from the model folder. That gives users common serving and chat entry points, but the engine underneath still determines hardware compatibility.
Edge0’s earlier 35B model uses a GGML-based native engine with a different platform matrix, including Apple Metal support. Those capabilities should not automatically be carried over to the new 145B and 120B releases.
There is a meaningful documentation gap here. The announcement promotes laptop operation and names a MacBook Pro as an example. GitHub also includes an Apple M2 Ultra benchmark row, yet its current support table marks macOS unavailable for the new models. A published performance row is evidence of a reported measurement, not necessarily a reproducible public setup.
For a developer choosing a machine today, I’d follow the explicit Linux/Windows NVIDIA path until Edge0 documents a supported alternative.
Q3_K and NVFP4 Need Separate Measurements
Quantization reduces the precision used to represent weights, lowering storage and data-transfer demands. Its practical effects also depend on kernel implementation, hardware and model quality after conversion. A lower-bit label does not guarantee a faster or better result on every device.
Edge0’s 145B benchmark table illustrates the point:
| Hardware | Format | 16K prefill | Chat decode | Agent decode |
|---|---|---|---|---|
| RTX 5090 D, 24GB | Q3_K | 3,554.9 tok/s | 99.7 tok/s | 121.2 tok/s |
| RTX 5090 D, 24GB | NVFP4 | 3,165.7 tok/s | 79.6 tok/s | 98.3 tok/s |
| RTX 4090, 24GB | Q3_K | 3,179.4 tok/s | 69.9 tok/s | 91.7 tok/s |
These are Edge0-reported measurements. The available sources provide no independent reproduction of the throughput, memory or capability claims.
On the listed RTX 5090 D configuration, Q3_K’s agent decode rate is about 23% higher than NVFP4’s, calculated from Edge0’s published figures. That makes Q3_K a sensible starting point for a speed evaluation on that hardware, but it says nothing by itself about whether the formats produce equivalent answers.
Prefill measures processing of input tokens. Decode measures generation of output tokens. Neither is a complete latency measurement: users also wait for model loading, prompt processing, tool execution and potentially many intermediate reasoning tokens.
The announcement’s 1,100-plus-token-per-second prefill and 80-plus-token-per-second decode claims belong to its laptop discussion. They should remain separate from the RTX 5090 D results unless the configurations and measurement methods are matched.
Frequently Asked Questions
3 questions
1Can Edge0-145B Run on a Mac Laptop Today?
Edge0 advertises laptop operation, but the current release documentation lists Linux and Windows with NVIDIA GPUs for the 145B and 120B models. It does not document a supported macOS path for them.
