Llama 4 (Scout and Maverick, both released April 2025) is Meta's best open-weights release yet, natively multimodal MoE architectures, with Scout's 10M-token context as the standout technical feat. The trade is honest: Llama 4 is behind the proprietary frontier (Claude, GPT, Gemini) on most coding and reasoning benchmarks, and the "Llama Community License" is custom rather than OSI-approved (commercial limits above 700M MAUs). The pick when self-hosting, on-prem, or per-token cost is the decisive factor.
- Scout 17B-active/109B-total MoE with 10M-token context
- Maverick 17B-active/400B-total MoE with 1M-token context
- Natively multimodal, text + image understanding
- Free weights, run on Bedrock, Together, Groq, Ollama, vLLM
- No API spend, no rate limits when self-hosted
- Behind the proprietary frontier on coding and reasoning benchmarks
- Custom "Llama Community License", not OSI-approved
- Commercial restrictions above 700M monthly active users
- Real infra cost, self-hosting MoEs is non-trivial at scale
- Behemoth (the larger sibling) delayed; the lineup gap matters
Llama is Meta’s open-weights large language model family. Where the proprietary frontier ships behind APIs you can’t audit, Llama ships the weights to download, runnable on your laptop via Ollama, on GPUs via vLLM, on managed inference via Bedrock, Together, Groq, and dozens of others. The 4.x line is Meta’s first natively multimodal release.
Where it fits
Llama is the default pick when self-hosting is non-negotiable, regulated industries, air-gapped deployments, sovereign-data requirements, or pure cost optimization at high volume. It’s also the substrate of choice for fine-tuning shops who need to add domain-specific capability without sharing data with a frontier vendor.
For pure benchmark performance, Llama 4 sits below the proprietary frontier on coding and reasoning. The wedge isn’t “best model”, it’s “best model you fully control.” For teams whose constraint is audit, governance, or cost per million tokens at volume, that’s a decisive wedge.
Cost to adopt
The weights are free under Meta’s Llama Community License. The real cost is inference infrastructure: GPU rental on cloud (AWS, GCP, Azure), managed inference providers (Together AI, Groq, Fireworks, Replicate), or self-hosted hardware. Typical managed pricing on Llama 4 Maverick lands at $0.20–0.60 per 1M input tokens, roughly 3–10× cheaper than proprietary frontier models. Self-hosted on your own GPUs can drive that to near-zero variable cost at the price of operational complexity. The Llama Community License does add a commercial-use restriction: services with more than 700M monthly active users need a separate Meta license.
How it compares
DeepSeek, Closest open-weights peer. V4 Pro is MIT-licensed (no commercial restriction) and scoring higher on most public benchmarks. Pick when license cleanliness or absolute performance matters.
Claude, Frontier proprietary, ahead on coding. Pick when API access and benchmark performance beat self-hosting.
GPT, Frontier proprietary, broadest ecosystem. Pick when ecosystem reach and multimodal generation matter.
Gemini, Frontier proprietary with the largest context. Pick when extreme context or video matters and self-hosting doesn’t.
What changed recently
Llama 4 Scout and Llama 4 Maverick released April 5, 2025, both natively multimodal MoE architectures, both shipping under the Llama Community License. Scout’s 10M-token context window remains the largest of any released model (proprietary or open). Through late 2025 and into 2026, third-party providers (Groq, Together, Fireworks, Replicate) brought managed inference pricing to roughly $0.20–0.60 per 1M input tokens for Maverick. Meta’s announced larger sibling, “Behemoth,” remains delayed, the gap in the top of the lineup has been a point of community criticism through 2026.
Sources
- Llama 4 herd announcement, ai.meta.com, Apr 5 2025
- Llama 4 models page, llama.com
- Welcome Llama 4 on Hugging Face, huggingface.co
- Llama 4 Open-Weights Specs, royfactory.net
- Llama 4: Did Meta just push the panic button?, interconnects.ai