Edge0

Edge0 is an Apache-2.0 local inference project for running its sparse models through a command line, local server, or chat application. On November 13, 2026, the project released Edge0-145B and Edge0-120B, latent mixture-of-experts checkpoints derived from Qwen3.8-Flash-Next. The release targets Linux or Windows systems with a supported NVIDIA GPU and CUDA 13; it is not a general Mac or phone deployment path for these newest weights.

What it provides

• Large local weights: 145B and 120B total-parameter checkpoints are offered in Q3_K and NVFP4 variants on Hugging Face. The project describes them as natively 4-bit-trained latent MoE models for local coding and agent work.
• One local interface: edge0-infer exposes chat and serve, while the project also documents a local OpenAI-compatible server. The native engine supports the older 35B model across more device classes.
• Explicit hardware boundary: the current 145B and 120B path uses edge0-strata on Linux or Windows with NVIDIA CUDA. Do not infer support for Apple Silicon from the older 35B engine's platform list.
• Private execution option: weights and serving can stay on hardware you control. That removes a hosted-model request from the workflow, but does not remove the need to secure the local endpoint, model files, prompts, and logs.

Practical workflow

1. Read the official model guide and confirm CUDA 13 plus the listed GPU and memory requirements.
2. Download one official 145B or 120B Hugging Face variant and install the matching Edge0 release.
3. Start with edge0-infer chat for a small coding or reasoning evaluation, then use edge0-infer serve only when a local application needs an API endpoint.
4. Keep the service bound to a trusted interface and test quality, latency, memory use, and license obligations on your real prompts before replacing a cloud model.

Limits and comparison

The project's published benchmark and throughput figures are useful starting points, but they are vendor-reported and hardware-specific rather than an independent guarantee. A 145B label also does not mean every workstation can run it: the newest models have a narrower NVIDIA/CUDA path than the older 35B engine. Teams that need a broadly supported local model can compare Qwen3.8-Flash-Next or choose a smaller checkpoint. Teams that need a managed API should not assume a local server has the same operational controls as a hosted provider.

FAQ

Is Edge0 open source?

The repository is licensed Apache-2.0. Check each downloaded model card and release artifact before redistribution because code and weights can have separate terms.

Can the current 145B model run on macOS?

Not according to the project's November 13 platform table. It lists the 145B and 120B models for Linux or Windows with an NVIDIA GPU and CUDA 13; macOS support there applies to the older 35B native-engine path.

Why use it instead of a hosted model?

It is a fit when local control, offline-capable operation, or avoiding per-request provider exposure matters and the hardware/operations tradeoff is acceptable.

Conclusion

Edge0 is a credible, actively maintained local-model option for builders who can meet its current hardware requirements. Treat the latest 145B and 120B release as a focused NVIDIA deployment route, validate it on your workload, and use the official repository as the source of truth for rapidly changing setup details.