Run Qwen3.5-122B on-premises
● Measured on our fleet — we run this model in production
Qwen3.5-122B-A10B is a 122-billion-parameter mixture-of-experts model that activates about 10 billion parameters per token — flagship-class reasoning at the serving cost of a mid-size model. It is one of the two models on this site we do not have to estimate: we run it in production, on the exact class of hardware we sell.
On our fleet it serves as an internal planning and review brain behind our own products, on a single GB10-class unit.
Memory fit on a 128 GB unit — measured
| Parameters | 122B total, ~10B active per token (mixture-of-experts) |
|---|---|
| Weights on device (4-bit NVFP4) — measured | 78 GB |
| Device memory | 128 GB unified (one unit) |
| Left for context cache + system | ~40 GB |
| Units required | 1 |
The 78 GB figure is not a datasheet estimate — it is the on-device footprint of the checkpoint we serve, from our fleet records. Because only ~10B parameters are active per token, the model stays responsive on a single unit.
What it serves well
- Planning and multi-step reasoning for agent workflows — the role it holds in our own fleet.
- Review and quality-judgment duty: assessing drafts, plans, and code changes.
- Long-form document analysis and synthesis on-premise.
- General enterprise-assistant duty at flagship quality, entirely on your hardware.
Our production deployment
This model runs daily in the BrainOutput fleet as an internal planner and reviewer, on the same class of machine we provision for customers. The deployment facts on this page — footprint, single-unit fit — come from our own fleet records, not from a vendor datasheet.
Honest limits
- We publish the memory facts we measured; we do not publish invented benchmark scores. The 'serves well' list above comes from daily production use.
- Tokens-per-second depends on your context lengths and concurrency — we size that in the assessment against your real workload, not against a synthetic number.
Frequently asked questions
- Does Qwen3.5-122B really fit on one unit?
- Yes — measured, not estimated: 78 GB of quantized weights on a 128 GB unified-memory unit, with roughly 40 GB left for the context cache and system. It is one of the models our own fleet serves daily.
- What do you use it for yourselves?
- Planning and review duty inside our own agent workflows: it drafts and assesses plans and acts as a quality judge for our other models' output. What we sell is the setup we run.
- Can it be fine-tuned on our data?
- Open-weight models in this class can be fine-tuned. Our LLM Factory prepares datasets, runs the fine-tune, benchmarks the result against the base model on your tasks, and delivers it onto the same device.