Meta

Key Terms

• Open-weight model - a model whose trained weights are published for anyone to download, modify, and deploy. Meta's Llama models are open-weight, meaning the cost of inference depends entirely on which hosting provider you choose. Source: Llama
• Inference provider - a third-party service (e.g., Cerebras, Together AI, Groq, AWS Bedrock, Azure AI) that hosts Llama models and charges for API access. Meta does not sell inference directly. Source: Llama
• Distributed inference - running inference across multiple GPUs or nodes to serve large models. Llama 4 Scout and Maverick both support distributed inference, with estimated costs of $0.19-$0.49/MTok. Source: Llama

Latest Changes

First report for this supplier. All models, plans, and pricing are listed as current state.

• New model: Llama 4 Scout and Maverick available as open-weight models. Both feature 10M token context windows and native multimodal support.
• Deprecation (upcoming): Cerebras deprecating Llama 3.1 8B on May 27, 2026.
• Plan change: Website migrated from llama.meta.com to Llama Old URL returns HTTP 400 errors.

Plans

Meta does not offer a coding agent product or subscription plans. Llama models are open-weight and free to download. The cost structure depends entirely on how you deploy them.

Source: Llama

API Pricing

Meta does not offer an API directly. Inference costs are set by third-party providers. Concrete pricing from Google Vertex AI and Cerebras:

Terms explained:

• Batch API - a lower-cost inference mode where requests are queued and processed asynchronously (not real-time), typically at 50% of standard pricing. Google Vertex AI offers batch pricing for all Llama models listed above. Source: Google – Pricing

Source: Llama Google – Pricing, Cerebras – Pricing

Model Performance / Benchmarks

Additional specifications:

• Both models: 10M token context window, text + image (natively multimodal)
• Llama 4 Scout: optimized for efficient inference on a single H100 GPU
• Llama 4 Maverick: targets frontier-level performance with higher resource requirements
• Estimated distributed inference cost: $0.19-$0.49/MTok

Source: Llama

Latest News

Llama 4 Scout and Maverick Release

Llama 4 Scout and Llama 4 Maverick are available as open-weight models. Both feature 10M token context windows and native multimodal support (text + image). Llama 4 Maverick achieves 80.5 MMLU Pro and 43.4 LiveCodeBench. Llama 4 Scout achieves 74.3 MMLU Pro and 32.8 LiveCodeBench. Both can be downloaded from Llama and are supported by major inference providers including Cerebras, Google Vertex AI, Together AI, Fireworks AI, AWS Bedrock, and Azure AI.

Cerebras Deprecation of Llama 3.1 8B

Cerebras is deprecating Llama 3.1 8B on its platform effective May 27, 2026. The model was priced at $0.10/MTok input and $0.10/MTok output. Users should migrate to Llama 4 Scout or Maverick before the deprecation date.

Website Migration

Meta's Llama website has migrated from llama.meta.com to Llama The old URL returns HTTP 400 errors. All model downloads and documentation are now hosted at the new domain.

Source: Llama Cerebras – Pricing

Community Signals

LiveCodeBench and Context Window Discussion

Llama 4 Maverick's LiveCodeBench score of 43.4 is frequently cited in coding benchmark discussions, with community members comparing it favorably to GPT-4-class models for code generation tasks. The 10M token context window is a major talking point, with developers noting it enables processing very large codebases in a single prompt. However, practical latency and cost at that context length are still being evaluated by the community.

Self-Hosting for Privacy

The lack of a first-party coding agent product from Meta means users must build their own agent infrastructure using tools like Continue, Cline, Aider, or OpenHands on top of Llama models. Organizations with strong privacy requirements (defense, healthcare, finance) often choose Llama models for on-premises deployment to avoid sending code to third-party APIs.

Provider Lock-In Concerns

The Cerebras Llama 3.1 8B deprecation has prompted discussion about provider lock-in when relying on a single inference provider for open-weight models.

Source: Llama

Enterprise Readiness

Transparency Gaps