Llama 4 Scout 17B and Llama 4 Maverick 17B arrived as fully managed, serverless options on Amazon Bedrock. These models occupy the U.S. East (N. Virginia) and U.S. West (Oregon) AWS Regions. Developers integrate these foundation models into applications through the Amazon Bedrock Converse API. This availability follows the introduction of the Llama 4 family last month. The models support diverse use cases such as intelligent agents, multilingual chat assistants, and image analysis for customer support.
Model intelligence and MoE architecture
Llama 4 Scout uses 17 billion active parameters and 16 experts. It fits on a single NVIDIA H100 GPU with Int4 quantization. This model delivers a 10M context window. It outperforms Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across many benchmarks. Llama 4 Maverick uses 17 billion active parameters and 128 experts. It contains 400 billion total parameters. Maverick achieves results comparable to DeepSeek v3 on reasoning and coding. It also beats GPT-4o and Gemini 2.0 Flash on several benchmarks. These models use distillation from Llama 4 Behemoth. Behemoth uses 288 billion active parameters and 16 experts. It outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks such as MATH-500 and GPQA Diamond. The training process for Llama 4 uses the MetaP technique to set hyper-parameters like per-layer learning rates and initialization scales. The training mixture exceeded 30 trillion tokens. This amount is more than double the Llama 3 pre-training mixture. The process used FP8 precision and 32K GPUs to achieve 390 TFLOPs per GPU. The language support includes 200 languages, and over 100 languages contain more than 1 billion tokens each. The architecture uses mixture-of-experts layers. A single token activates only a subset of the total parameters. This design improves inference efficiency and lowers model serving costs. Llama 4 Maverick uses 128 routed experts and one shared expert. Which models will be the first to integrate these MoE efficiencies into local hardware?
Bedrock pricing tiers and BDA automation
Amazon Bedrock offers six inference options. Standard pricing charges for input and output tokens used during processing. Flex mode costs 50% less than Standard for workloads that do not require immediate responses. Priority mode costs 75% more than Standard to put requests at the front of the queue. Batch mode reduces costs by 50% for large-scale inference tasks. You already know that token costs drive the budget, so focus on the multiplier effects. Prompt caching reduces input costs to approximately 10% of the input price for cached prefixes. Writing to the cache costs 1.25 times the input price.
Bedrock Data Automation uses per-page pricing. A 1,000 page document with 15 fields costs $40.00 at a rate of $0.040 per page. A 1,000 page document with 40 fields costs $45.00 at a rate of $0.045 per page. If a document has 40 pages and 20 fields, the cost is $1.60. If a document has 10 pages and 40 fields, the cost is $0.45. The total for both is $2.05. Video processing costs $0.050 per minute for standard output. A 60 minute video costs $3.00. Meeting audio processing costs $0.006 per minute. Processing 15,000 minutes of audio results in a $90.00 charge. Bedrock Knowledge Bases integration with Bedrock Data Automation uses standard output at $0.010 per page. A 1,000 page document costs $10.00.
Scaling and the economics of on-premise
Companies decide between cloud and on-premise based on usage volume and regulatory needs. Enterprises managing 70B-class workloads with 100 million tokens per month find that Bedrock costs significantly less than on-premise builds because the high capital expenditure for dedicated hardware exceeds the marginal cost of cloud tokens at this specific scale. For workloads with 2 billion tokens per month, on-premise builds provide a 2x cost advantage over the cheapest hosted alternatives. The break-even point for 70B-class inference against the cheapest open-model hosted API stays between 1.0 billion and 1.5 billion tokens per month on a 4x 5090 setup. Cloud GPU rental is the worst of both worlds. Rental rates for the box and the operational costs to run it make it less efficient than either cloud APIs or on-premise builds.
| Model Name | Active Parameters | Total Parameters | Context Length |
|---|---|---|---|
| Llama 4 Scout | 17B | N/A | 10M |
| Llama 4 Maverick | 17B | 400B | N/A |
| Llama 4 Behemoth | 288B | N/A | N/A |
Data security and sovereignty
Industries like healthcare, finance, and government use on-premise AI to maintain data sovereignty. This configuration keeps all data, models, and processing within internal infrastructure. Data sovereignty becomes absolute because queries and outputs stay within the network boundary. There are no API calls that leave the firewall. In contrast, cloud AI introduces risks from multi-tenancy. Workloads on shared infrastructure face side-channel attack vectors. Cloud users operate under a shared responsibility model. The provider secures the infrastructure, while the user secures the application layer. Data residency also presents challenges for compliance with GDPR. Cloud providers may replicate data across different geographic regions. Users also face risks from vendor access and potential API interception.
Llama 4 technical specifications
The Llama 4 architecture uses a mixture-of-experts (MoE) design. This design ensures that a single token activates only a fraction of the total parameters. Llama 4 Maverick has 17 billion active parameters and 128 routed experts. It also includes a shared expert. During inference, a token enters the shared expert and one of the 128 routed experts. This process improves inference efficiency and lowers serving costs. Maverick fits on a single NVIDIA H100 DGX host. The models use native multimodality. This design integrates text and vision tokens into a unified backbone through early fusion. The training process for Llama 4 includes specialized datasets for long context extension. This capability allows for extended context processing in applications.
| Model | Parameter Count | Architecture | Training Data |
|---|---|---|---|
| Llama 4 Scout | 17B (Active) | MoE | 30+ Trillion Tokens |
| Llama 4 Maverick | 400B (Total) | MoE | 30+ Trillion Tokens |
| Llama 4 Behemoth | 288B (Active) | MoE | 30+ Trillion Tokens |
Hidden costs and optimization
The cost of cloud AI involves more than just token counts. Egress charges ranging from $0.05 to $0.09 per GB apply when users ship images or video back from the service. Prompt caching reduces input costs by up to 90% for repeated prefixes. However, many cloud services log queries for performance monitoring. This behavior means sensitive prompts remain on vendor infrastructure. Users pay a premium of 1.25 times the input price to write to the cache.
To reduce costs, users should craft concise and clear prompts to eliminate unnecessary words. Precise language reduces ambiguity. Setting maximum token limits for model output helps control length. Developers can use libraries like Hugging Face Tokenizers to estimate token counts before sending prompts to a model. Batch mode is a significant cost lever. This mode costs 50% less than Standard for selected models. It is useful for large volumes of data that do not require immediate responses.
Deployment decision
The selection between cloud and on-premise deployment hinges on token volume and regulatory requirements. Organizations with 100 million tokens per month should use Amazon Bedrock. Organizations with 2 billion tokens per month should deploy on-premise hardware. Regulated industries must use on-premise systems to satisfy data sovereignty laws. Users with low-latency requirements for robotics should avoid cloud APIs.
