Run Llama 4 Scout on-premises
○ Memory-fit arithmetic — checkable math, not a benchmark
Llama 4 Scout is Meta's compact Llama 4 variant: 109 billion total parameters with 17 billion active (16 experts), natively multimodal and advertised with a very long context window. It fits a 128 GB machine at 4-bit with room for context.
Here is the arithmetic against a machine with 128 GB of memory.
Memory fit in 128 GB
| Parameters | 109B total, 17B active (16 experts, MoE) |
|---|---|
| 4-bit weights | ~55–60 GB |
| Reference memory | 128 GB (one machine) |
| Left for context cache + system | roughly 50 GB |
| Fits in 128 GB | yes |
At 4-bit Scout fits a 128 GB machine; its long advertised context is bounded in practice by the KV-cache memory you leave free, which we size in the assessment.
What it serves well
- Multimodal workloads (text + images) served entirely on-premise.
- Long-document tasks where a large context window helps.
- Teams standardizing on the Llama ecosystem and tooling.
Honest limits
- Memory arithmetic, not a benchmark.
- Llama 4 ships under Meta's Llama 4 Community License, which carries use restrictions (including a large-user clause). We review licensing fit as part of the assessment.
- The bigger Llama 4 Maverick (~400B) does NOT fit 128 GB — it needs a GPU server or a private cloud, or is out of scope.
Frequently asked questions
Does Llama 4 Scout fit in 128 GB of memory?
Yes, at 4-bit: ~55–60 GB of weights against 128 GB, leaving roughly 50 GB for context and system.
Is the license fully open?
No — it is the Llama 4 Community License with use restrictions, not a plain OSI license. We flag this rather than imply otherwise.
What about Llama 4 Maverick?
Maverick (~400B) does not fit in 128 GB of memory; it would require a GPU server or a private cloud. We say so plainly.
Find out which model fits your work
An AI diagnostic maps your processes, classes what is sensitive and sizes the model and the hardware against your real tasks.