Edge0 · 2026-11-12 · major
Edge0-145B — an open 145B MoE model that runs on one consumer GPU
In short
Edge0-145B is an open-weight 145B Mixture-of-Experts model with about 10B active parameters, built on Qwen3.8-Flash-Next. Edge0 reports 29 GB peak memory and 80+ tokens/s on a single consumer NVIDIA GPU.
Edge0 · 2026-11-12 · 4 sources checked
Edge0 open-sources a 145B coding and agent model that fits a single consumer GPU.
Key specs
| Terminal bench 2.1 | 86.5 |
|---|---|
| Peak memory | 29 GB |
Quick facts
| Maker | Edge0 |
|---|---|
| Parameters | 145B total, ~10.12B active per token |
| Base model | Qwen3.8-Flash-Next |
| Context | 256K tokens |
| Builds | Q3_K and NVFP4, plus Edge0-120B |
| License | Edge0 Community Model License 1.0 + Qwen license; framework Apache-2.0 |
| Runs on | Linux / Windows with an NVIDIA GPU (CUDA 13) |
Benchmarks
| Edge0-145B | 86.5% | |
|---|---|---|
| Qwen3.8-Flash-Next | 83.3% | |
| DeepSeek-V4-Flash-0731 | 82.7% | |
| Claude Opus 4.6 | 78.2% |
What is it?
Edge0-145B is a 145-billion-parameter open-weight model that Edge0 built to run on one computer instead of a data center. Released on 12 November 2026 together with a smaller Edge0-120B, it ships with Edge0's inference framework and an Edge0 App for running it without a build setup.
How does it work?
The model is a latent Mixture-of-Experts: each of its 48 layers has 512 experts, and only 10 routed experts plus 1 shared expert work on each token, so about 10.12B parameters are active. The experts use a narrow internal width, which makes each one a small record that the engine can keep on the GPU, hold in system RAM (about 40 GB) or swap in when needed, while about 48 GB of vocabulary data is read from SSD on demand.
Why does it matter?
Edge0's own tests put the 145B model ahead of Qwen3.8-Flash-Next and DeepSeek-V4-Flash on DeepSWE 1.1 and Terminal-Bench 2.1, at 29 GB measured peak memory instead of an estimated 70 GB for a standard 4-bit Qwen3.8-Flash-Next. That brings a frontier-class coding model into reach of people with one high-end gaming GPU, with no API bill and no data leaving the machine.
Who is it for?
developers who want a strong coding model running locally
Frequently asked questions
- How much memory does Edge0-145B need?
- Edge0 reports 29 GB of measured peak memory for Edge0-145B. That number is the GPU-side peak: the launch post also keeps an expert library of about 40 GB in system memory and reads about 48 GB of vocabulary data from the SSD, so a machine needs plenty of RAM and disk on top of the GPU.
- Can Edge0-145B run on a Mac?
- Edge0-145B and Edge0-120B run on the edge0-strata engine, which the Edge0 README lists for Linux and Windows with an NVIDIA GPU using CUDA 13. The README marks macOS as not supported for these two models today, even though Edge0 shows Apple M2 Ultra benchmark rows; the older Edge0-35B runs on Metal.
- Is Edge0-145B open source?
- Edge0's inference framework is Apache-2.0, but the Edge0-145B weights use the Edge0 Community Model License 1.0, and the Qwen Community License also applies because the model is built on Qwen3.8-Flash-Next. Hacker News commenters argued the model is open-weight rather than open source, so read both licenses before commercial use.
- How does Edge0-145B compare with Qwen3.8-Flash-Next?
- In Edge0's internal tests, Edge0-145B averages 72.8 across six benchmarks versus 71.7 for Qwen3.8-Flash-Next, and it leads on DeepSWE 1.1 (64.6 vs 58.7) and Terminal-Bench 2.1 (86.5 vs 83.3). Qwen3.8-Flash-Next scores slightly higher on GPQA-Diamond and HLE. These are Edge0's own numbers, not independent results.
Try it
build/cpu/bin/edge0-infer serve --model /path/to/Edge0-145B-q3k