Llama 4 is Meta's open-weight, multimodal (text and image) model family: Scout (17B activated/109B total, 10M-token context) and Maverick (17B activated/400B total, 1M-token context), both released April 2025 under Meta's own Community License, not Apache or MIT. Pick a specific variant, not "Llama 4" in general; the two differ enormously in size and hosting requirements.
Other good fits
Teams comfortable with Meta's own Community License terms rather than Apache-2.0 or MIT.
+ Pros
Read pros
"Llama 4" names a family, not one model: Meta released two variants on 5 April 2025. Scout has 17B activated parameters out of 109B total, a 10-million-token context window, and was trained on about 40 trillion tokens. Maverick has the same 17B activated parameters but 400B total, a 1-million-token context, and was trained on about 22 trillion tokens. Both use a mixture-of-experts architecture, are natively multimodal (text and image input, multilingual text/code output), and support 12 languages including English, German, French, Spanish, Arabic and Hindi.
Pick it when…
You need Scout's very long context (10M tokens) for a task that genuinely needs to reference a huge amount of text at once.
You need Maverick's larger total capacity and can host its 400B total parameters.
Native multimodal input (text plus images) matters for your task, and you can accept Meta's Community License.
− Cons
Read cons
Scout and Maverick differ so much in total parameters and context that neither one deployment plan fits both; don't treat "Llama 4" as a single specification. Meta's benchmark numbers and stated context length are the publisher's own claims, not a substitute for testing your specific multimodal workflow, image types and prompts.
Pick something else when…
You need a small text-and-tool-use model for modest local hardware and don't need image input.MiniCPM5-2B is a compact model built for that case instead.
You need an Apache/MIT-licensed open model. Check a model with permissive licensing terms rather than Meta's Community License if that's a hard requirement.
You want to serve it yourself efficiently. Use vLLM as the serving engine once you've picked a variant and secured the hardware.
Avoid if
You need Apache-2.0 or MIT licensing specifically; Llama 4 uses Meta's own commercial license terms.
Context
Scout and Maverick differ enormously: "Llama 4" alone doesn't specify a deployment. Pick the exact variant your project needs before sizing hardware or estimating cost, and check the current model card rather than assuming last year's numbers, since Meta continues to release updates to this family.
Requirements
Meta's published evaluations use BF16 precision; quantized checkpoints are also available for lower-memory deployment. Neither variant is a small local model: Scout's 109B total parameters and Maverick's 400B total parameters both need serious hardware or a hosted inference provider. This review didn't measure a specific hardware, precision and workload combination; test the exact variant and serving path you plan to use.
Licence & cost
The weights are free to download under the Llama 4 Community License Agreement, Meta's own terms rather than Apache-2.0 or MIT, so check its deployment conditions (including any user-count threshold that changes your obligations) before commercial use. There's no official workload cost published for either variant; expect substantial compute, storage and serving costs given the total parameter counts, or use a hosted inference provider and pay per token instead of running it yourself.
First task
Documentation-based suggestion — not run here.
Test multimodal understanding on a real image.
Pick the variant that fits your context and hardware needs (Scout for very long context, Maverick for larger capacity).
Give it a photo (a whiteboard, a product, a document) along with a short multilingual instruction.
Ask for a table separating what's directly visible in the image from what it's inferring or guessing.
Check each row against the actual image yourself.
Done when every "visible" row is something you can independently confirm is actually in the image, and every inferred row is clearly labelled as such.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Meta's Llama 4 model card, license text and cookbook repository, checked 23 September 2026. No model was downloaded, deployed or run for this review.
Llama 4 model card · Sep 23, 2026 — Scout and Maverick parameter counts, context lengths, training tokens, release date, multilingual support.
Llama 4 Community License · Sep 23, 2026 — Meta's own license terms, distinct from Apache-2.0 or MIT.
MiniCPM5-2B wins when you need a small, tool-capable text model on genuinely modest hardware instead of a large mixture-of-experts model, and don't need image input.
vLLM wins as the serving layer once you've chosen a Llama 4 variant and have the hardware to run it.
Ollama wins when you want the easiest local-model experience and can use a smaller quantized model instead of Scout or Maverick's full size.
Run a multimodal model for broad tasks with a large context window.
Use quantization to target a single high-end GPU.
Plan deployments around Meta's Community Licence conditions.