> ## Content Index
> Fetch the complete content index at: https://claudefox.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Edge0-145B: pitched for laptops, but no laptop appears in its speed table
- URL: https://claudefox.com/edge0-145b-open-model-local-hardware/
- Published: 2026-11-13T10:09:35.000Z
- Updated: 2026-11-20T03:29:33.000Z
- Description: Edge0 released Edge0-145B and Edge0-120B, sparse models derived from Qwen3.8-Flash-Next that keep hot experts on the GPU and stream the rest from RAM and disk. The launch post talks about laptops; the speed tables list RTX cards, Spark and an M2 Ultra. Setup, sizes, licence and caveats.
- Author: hupicake
- Tags: News

*Updated November 20, 2026: Updated: the repository now lists Apple Silicon (Metal) support for both models. Added setup commands, memory measurements, 120B speeds and licence terms.*

Edge0 released two models in mid-November, Edge0-145B and Edge0-120B, along with its inference software (the launch post is dated November 12, the README's news entry November 13). The 145B model activates about 10.1 billion parameters per token and the 120B about 8.6 billion. The repository's model page says both are derived from Qwen3.8-Flash-Next. Edge0 calls the release open-source; the framework is Apache 2.0, but the weights carry a community licence with commercial limits (see below). The launch post calls Edge0-145B the first open on-device model at that size, and says it runs on a compatible personal laptop such as a MacBook Pro. The measurements Edge0 publishes tell a more specific story.

## How it fits in less memory

Both models are mixtures of experts: 48 layers, each with 512 routed experts plus one shared expert, and each token uses 10 of them. Edge0 calls the design latent MoE. Expert inputs are compressed to a narrower width, 1,792 for the 145B and 1,280 for the 120B, while attention stays at full width. Every expert therefore becomes a small fixed-size record. The engine keeps a few thousand of those records on the GPU and streams the rest from system memory. A 51.2-billion-parameter n-gram embedding table is memory-mapped from disk. Edge0 says the models were trained natively at 4 bits, and a multi-token prediction head drives speculative decoding.

## What the launch post claims, and what the tables show

The launch post reports 29GB peak memory (elsewhere in the same post, 28GB), more than 1,100 tokens per second of prompt processing, more than 80 tokens per second of generation and about 120W typical power. It does not say which machine produced those numbers. The repository's speed table for Edge0-145B lists these chat-decode rates:

- RTX 5090 D 24GB: 99.7 tokens per second with the Q3\_K build, 79.6 with NVFP4.
- RTX 4090 24GB: 69.9 tokens per second (Q3\_K).
- Apple M2 Ultra, using Metal: 60.9 tokens per second (Q3\_K).
- RTX Spark N1X with 60GiB: 37.7 (Q3\_K) and 40.1 (NVFP4).
- The same RTX 5090 D limited to 16GiB of VRAM and 20GiB of RAM: 36.8\. The Hugging Face card adds a 12GiB VRAM test at 22.4.

Edge0 says the capped rows are test limits, not minimum requirements, and there is no capped result for the 145B NVFP4 build. With nothing capped, the 145B Q3\_K build occupied about 23GiB of VRAM plus 31GiB of system RAM on the 5090 D. Agent decode, the README's headline figure, runs higher: 121.2 tokens per second on the 5090 D. The smaller 120B is faster on the same hardware: 130.2 tokens per second chat decode on the 5090 D, 95.0 on the 4090, 60.4 on the M2 Ultra and 62.4 in the 16GiB-capped test. No laptop appears in either table.

The README now lists Metal on Apple Silicon as a backend for both models, and the Edge0 Chat desktop app for macOS can download and run all four model folders. On macOS the Metal template uses a 32,768-token context (the model page lists 256K overall), and the first start writes a file into the model folder, which must be writable.

## How to run it

Weights are on Hugging Face in four folders: Edge0-145B-q3k, Edge0-145B-NVFP4, Edge0-120B-q3k and Edge0-120B-NVFP4\. The build weights are 40.81GiB (Q3\_K) and 47.79GiB (NVFP4) for the 145B, and 27.98GiB and 32.96GiB for the 120B. The full NVFP4 repository on Hugging Face is about 114GB, so plan disk space accordingly. If you would rather not compile anything, the Edge0 Chat app for Windows and macOS on the GitHub releases page is the alternative. From the command line, you clone the repository, build the edge0-strata engine and edge0-infer following their docs, then serve a model folder:

```
git clone https://github.com/Edge0-AI/Edge0.git edge0
# build edge0-strata and edge0-infer per their docs, then from edge0-infer/:
build/cpu/bin/edge0-infer serve --model /path/to/Edge0-145B-q3k

curl localhost:8000/v1/chat/completions -H 'content-type: application/json' \
  -d '{"model":"edge0_145b","messages":[{"role":"user","content":"hi"}],"max_tokens":64}'
```

The server listens on 127.0.0.1:8000 and exposes both OpenAI-compatible and Anthropic-compatible endpoints; we have not tested specific tools against it. On Linux and Windows the engine needs an NVIDIA GPU with CUDA 13\. For tight memory, the README documents an SSD mode (--mmap-experts) and a --resident-budget-gib cap on the RAM expert cache.

## Benchmarks and licence

The scores are Edge0's own tests. It reports 86.5 on Terminal-Bench 2.1 for the 145B model, against 83.3 for Qwen3.8-Flash-Next, and 64.6 on DeepSWE 1.1 against 58.7\. The figures for Edge0-120B still don't match between Edge0's own pages. The launch post gives 81.3 on Terminal-Bench 2.1 and 53.6 on DeepSWE 1.1\. The README and model page give 75.3 and 51.3\. Wait for independent tests before relying on any of them.

The inference framework is Apache 2.0, but the weights use the Edge0 Community Model License 1.0, and Qwen's own licence still applies to the Qwen-derived parts. According to the model card, personal use, research, internal use and commercial use of generated outputs are allowed. Selling or bundling the model, running a commercial model service, or offering a commercial coding or office assistant built on it needs a separate written licence from Edge0.

## Related on claudefox

- [llama-modchip runs the 284B DeepSeek-V4-Flash on a 64 GB PC with no GPU, using patched llama.cpp](https://claudefox.com/llama-modchip-deepseek-v4-flash-64gb-cpu/)
- [Researcher says released weights show six Chinese models were started on Nvidia Blackwell GPUs](https://claudefox.com/checkpoint-forensics-chinese-models-blackwell-fingerprint/)
- [Hugging Face Inference Providers now route through OpenRouter: what changes for your code, keys and Claude Code](https://claudefox.com/hugging-face-inference-providers-openrouter-routing-custom-keys/)
- [Qwen4-Max is out: 1M context, $1.50/$4.50 pricing, and a Claude Code setup](https://claudefox.com/qwen4-max-release-pricing-claude-code/)

## Our take

The engineering is well documented: compressed experts, tiered streaming and clear notes on how each measurement was taken. In our view the marketing runs ahead of it. The published speeds come from 24GB RTX cards, an RTX Spark system and an M2 Ultra, not a laptop, and in Edge0's own uncapped measurement the 145B Q3\_K build took about 54GiB of GPU and system memory combined. If you already have a 4090 or 5090 and want a local coding model behind an Anthropic-compatible endpoint, it may be worth trying, though we have not measured install time or tool compatibility. Check the licence before building anything you plan to sell.

Source: [Introducing Edge0-145B (Edge0 blog)](https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer?ref=claudefox.com), [Edge0 repository README (GitHub)](https://github.com/Edge0-AI/Edge0?ref=claudefox.com), [Edge0 145B and 120B model page (GitHub)](https://github.com/Edge0-AI/Edge0/blob/main/docs/models/edge0-145b.md?ref=claudefox.com), [Edge0-145B-NVFP4 model card (Hugging Face)](https://huggingface.co/Edge0/Edge0-145B-NVFP4?ref=claudefox.com)

This story first appeared in the [claudefox daily briefing for November 13, 2026](https://claudefox.com/daily-2026-11-13/).