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 Qwen3-235B-A22B on-premises
○ Memory-fit arithmetic — checkable math, not a deployment claim
Qwen3-235B-A22B is Qwen's flagship open-weight model — 235 billion total parameters with about 22 billion active per token across 128 experts, under the permissive Apache-2.0 license, with a native 262K-token context window. Its sparse design gives near-frontier capability at roughly a third of the compute of a dense model its size.
It is also the point where honesty matters most: 235B does not fit comfortably on a single 128 GB unit. Here is the arithmetic — and the two-unit configuration that actually serves it.
Memory fit on 128 GB units
| Parameters | 235B total, ~22B active per token (128 experts, MoE) |
|---|---|
| Full precision (FP16) | ~470 GB — several units |
| 8-bit weights | ~235 GB — two units, no working room |
| 4-bit weights | ~118 GB at a clean 4-bit; ~130–140 GB for common Q4_K_M quants |
| One 128 GB unit | impractical — a clean 4-bit only fills it, and typical Q4 quants overflow it |
| Units required | two stacked 128 GB units at 4-bit (our measured DeepSeek-V4-Flash configuration) |
The rule of thumb (parameters × bytes per parameter) puts even a clean 4-bit at ~118 GB — filling a single unit with nothing left for context — and common Q4_K_M quants land ~130–140 GB, past what one unit holds. The honest answer is two stacked units, the exact hardware envelope we run measured every day. A ~3-bit quant (~100–110 GB) can squeeze onto one unit at a real quality cost.
What it serves well
- Near-frontier open-weight capability kept entirely on hardware you own.
- Long-context analysis (262K native) over large private corpora.
- Efficient serving for its class — ~22B active parameters, roughly a third of a dense 235B's compute.
- The demanding tier of a mixed fleet, sitting above the single-unit models on this hub.
Honest limits
- This is arithmetic, not a Qwen3-235B deployment of ours — but the two-stacked-unit envelope it needs is exactly the one we run measured with DeepSeek-V4-Flash (~149 GB FP8 across two units), so the sizing rests on hardware we operate daily.
- It is not a single-unit model: anyone who tells you a 235B runs comfortably on one 128 GB box is rounding away the context cache. We will not.
- A ~3-bit single-unit quant is possible but trades accuracy for fit; we would only recommend it after measuring the quality cost for your workload.
Frequently asked questions
- Does Qwen3-235B-A22B fit one 128 GB unit?
- Not comfortably. A clean 4-bit is ~118 GB — it fills the unit with no room for context — and common Q4_K_M quants (~130–140 GB) overflow it. The practical configuration is two stacked units at 4-bit; a ~3-bit quant can squeeze onto one unit at a quality cost.
- Is this the same hardware you run in production?
- Yes — two stacked 128 GB units is exactly our measured DeepSeek-V4-Flash configuration (~149 GB FP8, tensor-parallel across two units). The memory envelope on this page is one we operate daily.
- License?
- Apache-2.0 — permissive and commercial-friendly, no separate agreement required.