Llama 4 is worth approaching with clear eyes. Two models shipped with open weights and both are genuinely capable. A third was previewed and never released. And Meta has since moved its flagship line to closed models, which means Llama 4 is likely the end of an era rather than the middle of one.
None of that makes Scout and Maverick less useful. It does change how you should think about building on them. This guide covers what is actually available, what each model needs, and how to deploy on Runpod.
What is Llama 4?
Llama 4 is Meta's first natively Mixture-of-Experts generation, and the first Llama family built multimodal from the start. Two models are available with open weights:
- Llama 4 Scout: roughly 109B total parameters across 16 experts, with a 10 million token context window. The headline feature is that context length, which is far beyond anything else in general release.
- Llama 4 Maverick: roughly 400B total parameters across 128 experts, with a 1M token context window. The stronger of the two on general capability.
A third model, Behemoth, was previewed as a roughly 2T-parameter teacher model and never shipped. Meta reportedly hit routing and attention problems at that scale and lost confidence that the result justified release. There has been no formal cancellation, but it has not appeared, and it would be unwise to plan around it.
Both Scout and Maverick are available on Hugging Face and llama.com under Meta's community license. That license is more restrictive than Apache 2.0 or MIT, so read the terms before building commercially on either.
Why run Llama 4 on Runpod?
Scout's 10M context needs memory you cannot buy casually. The model weights are only part of the picture. KV cache at long context is where the real memory goes, and exercising that window properly means high-VRAM hardware.
MoE models are memory-heavy, not compute-heavy. Scout activates a fraction of its 109B parameters per token, but all of them have to be resident. That shifts the constraint from raw compute to VRAM capacity, which is exactly the tradeoff a rented GPU handles well.
Evaluation before commitment. Given that Meta has moved on from open weights, a sensible approach is to test whether Llama 4 does what you need before architecting around it. Renting makes that an inexpensive question to answer.
GPU requirements for Llama 4
Scout was designed to fit on a single high-end GPU at reasonable precision. Maverick was not.
- Scout, 4-bit quantized: roughly 60 GB. H100 (80 GB) from $1.99/hr on Community Cloud.
- Scout, bf16: roughly 220 GB. Multi-GPU, or a single B300 (288 GB) from $6.94/hr.
- Scout at long context: add substantial headroom for KV cache. H200 (141 GB) from $3.59/hr is a better starting point than an H100 if you intend to use the context window.
- Maverick, 4-bit quantized: roughly 220 GB. Multi-GPU or a single B300.
- Maverick, bf16: roughly 800 GB. Multi-node Clusters.
The practical advice: start with Scout quantized on an H100 or H200. It is the model Meta optimized for single-GPU deployment, and it is where the distinctive capability lives.
Step 1: Get access and create an account
Llama 4 weights are gated. Request access on Hugging Face or llama.com and accept Meta's license terms first, then generate a Hugging Face token, because the download will fail without one.
Then go to runpod.io and sign up with email, GitHub, or Google. Add credits or a payment method.
Step 2: Deploy
- Create a network volume of at least 300 GB and mount it at
/workspace. Llama 4 weights are large and you do not want to download them twice. - Go to the Pods section and select an H100 or H200 for Scout, or a multi-GPU configuration for Maverick.
- Attach your network volume and choose a vLLM or PyTorch template.
- Set your Hugging Face token as an environment variable so the gated download succeeds.
- Deploy.
Step 3: Serve and test
vLLM supports the Llama 4 architecture. Set max model length conservatively at first: Scout's 10M context is a capability, not a default. Loading at full context will consume enormous KV cache before you have sent a single request.
Start at 32k, confirm generation works, then raise context in steps while watching memory. This is the single most useful habit when working with long-context models.
Pro tips
Do not start at 10M context. It is the reason to choose Scout, but it is not the setting to debug on. Get the deployment working at ordinary context first.
Long context changes retrieval economics. If you have been building retrieval pipelines to work around context limits, Scout is worth testing against that architecture directly. Sometimes a large window is simpler than a good retriever.
Read the license properly. Meta's community license carries conditions that Apache 2.0 and MIT do not. If you are shipping commercially, have someone check it.
Keep weights on a network volume. At Llama 4 sizes, re-downloading is a meaningful cost in both time and money.
Stop your pod when you are done. Runpod bills by the second while a pod runs.
Wrapping up
Scout and Maverick remain strong open-weights models, and Scout's 10M context is still unusual enough to be worth building around if long-document work is your problem. Just go in knowing the family is not on an upward trajectory, and weigh that against alternatives like Qwen or GLM that are still actively shipping open weights.
Ready to try it? Deploy an H200 pod and start with Scout at moderate context.
FAQ
What is the difference between Llama 4 Scout and Maverick?
Scout is roughly 109B parameters across 16 experts with a 10M token context window, designed to fit on a single high-end GPU. Maverick is roughly 400B across 128 experts with a 1M context window and stronger general capability, but it needs multi-GPU deployment.
What happened to Llama 4 Behemoth?
It was previewed as a roughly 2T-parameter model and never publicly released. Meta reportedly encountered routing and attention issues at that scale. There has been no formal cancellation, but it has not shipped and should not be planned around.
What GPU do I need for Llama 4?
Scout quantized to 4-bit needs roughly 60 GB, so an H100 from $1.99/hr. If you intend to use long context, an H200 at $3.59/hr gives you room for KV cache. Maverick needs multi-GPU at any practical precision.
Is Llama 4 free for commercial use?
It ships under Meta's community license, which is more restrictive than Apache 2.0 or MIT and carries specific conditions. Review the terms before deploying commercially rather than assuming open weights means unrestricted use.
Is Meta still releasing open-weights Llama models?
Meta has moved its flagship line to closed models. Scout and Maverick remain available and supported, but the open-weights Llama roadmap is not what it was. If long-term open-weights availability matters to your architecture, factor that in.
