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 Qwen3-30B-A3B on-premises
○ Memory-fit arithmetic — checkable math, not a deployment claim
Qwen3-30B-A3B is a mixture-of-experts model with 30 billion total parameters and about 3 billion active per token, under the permissive Apache-2.0 license. Its small active footprint makes it fast and cheap to serve, and it leaves so much memory free that one box can host it alongside other models.
Here is the arithmetic on a GB10-class 128 GB unit.
Memory fit on a 128 GB unit
| Parameters | 30B total, ~3B active per token (MoE) |
|---|---|
| 4-bit weights | ~15–18 GB |
| Device memory | 128 GB unified (one unit) |
| Left for context + other models | roughly 100 GB |
| Units required | 1 (with room for more models) |
At ~15–18 GB the model uses a small fraction of the box, so the same unit can serve it plus an embedding model and a coder model concurrently.
What it serves well
- High-throughput chat, routing, extraction, and agent steps at low serving cost.
- Multi-model boxes where several small/mid models run side by side.
- Latency-sensitive workloads that don't need a 100B-class brain.
Honest limits
- Arithmetic, not our measured deployment: we run larger models in production, but the memory math here is straightforward and generous.
- A ~3B active budget is efficient, not maximal — for the hardest reasoning we would pair or route to a bigger model, which the same box can also hold.
Frequently asked questions
- Can one box run this plus other models?
- Yes — at ~15–18 GB it leaves roughly 100 GB free, enough to co-host embedding and coder models on the same unit.
- License?
- Apache-2.0 — permissive and commercial-friendly.
- Is it fast?
- Its ~3B active-parameter budget makes it markedly faster to serve than a dense model of similar total size; we quantify it for your workload in the assessment.