Runs on a private AI box you own
One compact GB10-class device — 128 GB unified memory — runs it on-premise, sold and provisioned by us. No cloud API keys, no data leaving the building.
Run Qwen3-Coder-30B on-premises
○ Memory-fit arithmetic — checkable math, not a deployment claim
Qwen3-Coder-30B-A3B is a sparse mixture-of-experts model built for code — 30.5 billion total parameters with about 3 billion active per token, under the permissive Apache-2.0 license, with a native 256K-token context window (extendable further with YaRN). The small active budget makes it fast; the long context makes it useful across a whole repository, not just a single file.
It is a natural fit for a private coding agent: pointed at OpenCode on the device, your source never leaves the machine. Here is the arithmetic on a 128 GB unit.
Memory fit on a 128 GB unit
| Parameters | 30.5B total, ~3B active per token (128 experts, MoE) |
|---|---|
| 4-bit weights (the practical choice) | ~15–18 GB |
| 8-bit weights | ~31 GB |
| Device memory | 128 GB unified (one unit) |
| Left for context cache + system at 4-bit | roughly 100+ GB — room for a repo-scale context |
| Native context | 256K tokens (extendable via YaRN) |
Weights are the small part here: at 4-bit the ~15–18 GB of parameters leave the overwhelming majority of a 128 GB unit for the KV cache — which is exactly what a long, repository-scale context needs. The ~3B active budget also keeps it fast to serve.
What it serves well
- A private coding agent that reasons over a whole repository, running on the device — pair it with OpenCode so source never leaves the machine.
- Agentic tool use and function calling for automated engineering workflows.
- Fine-tuning into a house coder on your own codebase via the LLM Factory, then delivered back onto the device.
- Fast interactive assistance thanks to the ~3B active-parameter budget.
Honest limits
- Arithmetic, not our measured deployment — though it sits well inside the envelope we run measured on this hardware, with room to spare.
- A ~3B active budget is efficient, not the raw capability of a frontier dense model; for the hardest reasoning we would size a larger model and quantify the trade-off in the assessment.
- Very long contexts consume the KV-cache headroom that makes this model attractive — real throughput at 256K tokens is workload-dependent; we measure it for your case.
Frequently asked questions
- Does Qwen3-Coder-30B fit one unit?
- Easily — ~15–18 GB of weights at 4-bit against 128 GB, leaving roughly 100 GB for the context cache. That headroom is the point: it is what lets the model hold a repository-scale context.
- License?
- Apache-2.0 — permissive and commercial-friendly, no separate agreement required.
- How does this pair with OpenCode?
- OpenCode runs on the device and points at the local model, so your code is analysed and generated entirely on hardware you own. Our own coding workers run OpenCode against our fleet the same way.