AI model variants: SmolVLM vs Qwen3

Every Everlock build embeds one vision-language model, and which one it carries decides the AI runtime's capabilities — how strong its answers and captions are, and how much memory and disk it needs. There are two variants, and you pick between them when you download or build (the model is baked in; there is no runtime switch).

Both variants are vision-capable and drive the same two features:

• the operator prompt inside the admin SSH session
• image captioning for the image backend

The only difference is which model does that work.

	🪶 SmolVLM (default)	🧠 Qwen3
Model	SmolVLM-256M-Instruct	Qwen3.5-2B
Size on disk	~250 MB	~1.9 GB
Answer & caption quality	basic	stronger
Memory & speed	low / fast	higher / slower
Best for	light hosts, the default choice	higher-quality output when you have the resources

SmolVLM — small and light (default)

The default build embeds SmolVLM-256M-Instruct. At ~250 MB it keeps the binary and container image light and runs comfortably on modest hardware, while still being vision-capable. It's the right choice for most deployments: basic operator help and serviceable image captions without a heavy footprint. If you don't have a specific reason to choose otherwise, use SmolVLM.

Qwen3 — larger and stronger

The Qwen3 build embeds Qwen3.5-2B. It produces noticeably stronger answers and captions, at the cost of a much larger download (~1.9 GB), more memory, and slower inference on CPU. Choose it when output quality matters more than footprint and the host has the resources to spare.

How to select a variant

The variant is fixed in the artifact you obtain — you choose it by which one you get:

• Download — the variant is a path segment in the URL: …/dl/latest/smolvlm/… or …/dl/latest/qwen3/….
• Container image — a separate tag per variant, e.g. cr.everlock.sh/everlock:0.5-smolvlm vs :0.5-qwen3 (the latest tag is the SmolVLM build).
• Build from source — SmolVLM is the default; add --features qwen3 for the larger model.

To switch later, download or build the other variant and apply it as an update — there is no flag that swaps the model in a running instance.

Read next

• The gguf-runner inference engine — how the models are embedded and the build/image matrix
• Deployment — where you pick the variant when installing
• AI runtime and access model