hidekazu-konishi.com
Amazon Bedrock Models as of 2024 - An Analysis of the Comprehensive Model Catalog
First Published:
Last Updated:
In this article, I will focus on the announcements related to Amazon Bedrock, summarize the overall framework of Amazon Bedrock announcements at AWS re:Invent 2024, and consider an overview of Amazon Bedrock models as of the end of 2024.
Summary of Overall Amazon Bedrock Announcements at AWS re:Invent 2024
First, I have categorized the Amazon Bedrock announcements from AWS re:Invent 2024 into six categories from my perspective.Within this framework, items marked with (*) are announcements related to Amazon Bedrock models.
- Enhancement of RAG (Retrieval Augmented Generation) functionality
- (*) Optimization of search with Rerank model and API introduction
- Improvement of user experience through streaming output
- Implementation of custom connectors and streaming ingestion
- Integration with GraphRAG and Kendra GenAI index
- Expansion of multimodal support
- (*) Provision of various models through Nova series (text, image, video, audio)
- Multimodal data ingestion for knowledge base
- Toxicity detection function for images
- Model optimization and performance improvement
- Introduction of latency optimization options
- Implementation of model distillation functionality
- Support for prompt caching
- Prompt routing functionality
- Ecosystem expansion
- (*) Introduction of Bedrock marketplace
- (*) Addition of new model providers
- (*) Integration of third-party models
- Strengthening of quality control and evaluation functions
- Model evaluation using LLM-as-a-judge
- Evaluation function for knowledge base
- Functional expansion of guardrails
- Introduction of automatic inference checks
- Promotion of automation and efficiency
- Introduction of data automation functionality
- Support for structured data queries
- Multi-agent collaboration functionality
- Improvement of reliability and safety assuming enterprise use
- Improvement of performance and efficiency of the entire Bedrock utilization system
- Improvement of user and developer experience and lowering of implementation barriers
- Support for a wider range of use cases
Number of Amazon Bedrock Models (as of 2024-12-16)
Next, I'll summarize the number of models available in Amazon Bedrock.In the Amazon Bedrock model catalog, models that can be used when enabled in Amazon Bedrock Model Access have come to be called Serverless models to distinguish them from models available from the Amazon Bedrock Marketplace.
The following calculation formula represents the number of models available in Amazon Bedrock in N. Virginia [us-east-1] and Oregon [us-west-2] as of 2024-12-16.
Serverless (us-east-1): |E| = 46
Serverless (us-west-2): |W| = 44
Serverless (us-east-1 ∪ us-west-2): |E ∪ W| = 52
Bedrock Marketplace: |M| = 122
Total: |E ∪ W| + |M| = 174
In other words, as of 2024-12-16, there are 52 types of Serverless models when merged, 122 types in the Bedrock Marketplace, and a total of 174 types of models available in N. Virginia [us-east-1] and Oregon [us-west-2].
List of Amazon Bedrock Serverless Models: By Region (as of 2024-12-16)
Next, I'd like to look at the list of 52 types of Serverless models mentioned in the calculation formula earlier, merged for N. Virginia [us-east-1] and Oregon [us-west-2] as of 2024-12-16, by region.Here's how to read the table:
* List of models by provider and model type (merged for N. Virginia [us-east-1] and Oregon [us-west-2])
* Italics (red) are models available only in N. Virginia [us-east-1]
* Bold (blue) are models available only in Oregon [us-west-2]
| Provider | Modality | Model Name |
|---|---|---|
| AI21 Labs | Text | Jurassic-2[Ultra, Mid], Jamba 1.5[Mini, Large], Jamba-Instruct |
| Amazon | Text | Nova Micro, Titan Text G1[Lite, Express, Premier], Rerank 1.0 |
| Amazon | Text & Vision | Nova[Pro, Lite] |
| Amazon | Image | Nova Canvas, Titan Image Generator G1[v1, v2] |
| Amazon | Video | Nova Reel |
| Amazon | Embeddings | Titan Multimodal Embeddings G1, Titan Text Embeddings V2, Titan Embeddings G1 – Text |
| Anthropic | Text | Claude 3.5[Haiku], Claude[v2.1, v2.0], Claude Instant[v1.2] |
| Anthropic | Text & Vision | Claude 3.5[Sonnet v2, Sonnet], Claude 3[Haiku, Sonnet, Opus] |
| Cohere | Text | Command R+, Command R, Command, Command Light, Rerank 3.5 |
| Cohere | Embeddings | Embed Multilingual, Embed English |
| Meta | Text | Llama 3.2[1B Instruct, 3B Instruct], Llama 3.1[8B Instruct, 70B Instruct, 405B Instruct], Llama 3[8B Instruct, 70B Instruct] |
| Meta | Text & Vision | Llama 3.2[90B Vision Instruct, 11B Vision Instruct] |
| Mistral AI | Text | Mistral Large 2, Mistral Large, Mistral Small, Mixtral 8x7B Instruct, Mistral 7B Instruct |
| Stability AI | Image | SD3 Large 1.0, Stable Image Core 1.0, Stable Image Ultra 1.0, SDXL(1.0) |
List of Amazon Bedrock Serverless Models: By Provision Period (as of 2024-12-16)
Next, let's look at the list of Serverless models merged for N. Virginia [us-east-1] and Oregon [us-west-2] from a different perspective, categorized by provision period.Here's how to read the table:
* List of models by provider and model type (merged for N. Virginia [us-east-1] and Oregon [us-west-2])
* Italics (green) are models that became available from January 2024 to before AWS re:Invent 2024
* Bold (brown)) are models that became available at AWS re:Invent 2024
| Provider | Modality | Model Name |
|---|---|---|
| AI21 Labs | Text | Jurassic-2[Ultra, Mid], Jamba 1.5[Mini, Large], Jamba-Instruct |
| Amazon | Text | Nova Micro, Titan Text G1[Lite, Express, Premier], Rerank 1.0 |
| Amazon | Text & Vision | Nova[Pro, Lite] |
| Amazon | Image | Nova Canvas, Titan Image Generator G1[v1, v2] |
| Amazon | Video | Nova Reel |
| Amazon | Embeddings | Titan Multimodal Embeddings G1, Titan Text Embeddings V2, Titan Embeddings G1 – Text |
| Anthropic | Text | Claude 3.5[Haiku], Claude[v2.1, v2.0], Claude Instant[v1.2] |
| Anthropic | Text & Vision | Claude 3.5[Sonnet v2, Sonnet], Claude 3[Haiku, Sonnet, Opus] |
| Cohere | Text | Command R+, Command R, Command, Command Light, Rerank 3.5 |
| Cohere | Embeddings | Embed Multilingual, Embed English |
| Meta | Text | Llama 3.2[1B Instruct, 3B Instruct], Llama 3.1[8B Instruct, 70B Instruct, 405B Instruct], Llama 3[8B Instruct, 70B Instruct] |
| Meta | Text & Vision | Llama 3.2[90B Vision Instruct, 11B Vision Instruct] |
| Mistral AI | Text | Mistral Large 2, Mistral Large, Mistral Small, Mixtral 8x7B Instruct, Mistral 7B Instruct |
| Stability AI | Image | SD3 Large 1.0, Stable Image Core 1.0, Stable Image Ultra 1.0, SDXL(1.0) |
This is natural as Amazon Bedrock became General Availability (GA) on 2023-09-28, but I can also see a trend where older versions of models provided at the initial GA are gradually phased out as new versions appear.
From this, it is important to identify which models the newly announced models correspond to as successors, and to understand the functions and features of new models, confirm their substitutability, and prepare for migration as early as possible, assuming the replacement of older model versions.
Serverless Models Announced at AWS re:Invent 2024 (Available)
Next, I'll introduce the new Serverless models announced as available at AWS re:Invent 2024.Although not listed in this table, among the models announced as coming soon at AWS re:Invent 2024, which will be introduced in the next heading, Stable Diffusion 3.5 Large became available on 2024-12-19.
Here's how to read the table:
* List of models by provider and model type (merged for N. Virginia [us-east-1] and Oregon [us-west-2])
* Italics (red) are models available only in N. Virginia [us-east-1]
* Bold (blue) are models available only in Oregon [us-west-2]
| Provider | Modality | Model Name | Overview |
|---|---|---|---|
| Cohere | Text | Rerank 3.5 | A model to improve search accuracy in RAG applications. It takes user queries and retrieved document sets as input, re-ranks them based on relevance, and prioritizes the selection of optimal documents as model input to improve the quality of generated responses. |
| Amazon | Text | Rerank 1.0 | A model to improve search accuracy in RAG applications. It takes user queries and retrieved document sets as input, re-ranks them based on relevance, and prioritizes the selection of optimal documents as model input to improve the quality of generated responses. |
| Amazon | Text | Nova Micro | A text-only model capable of high-speed processing with minimal latency. Optimized for basic tasks such as summarization, translation, classification, dialogue, and coding with a context length of 128K tokens. Supports fine-tuning. |
| Amazon | Text & Vision | Nova Lite | A low-cost multimodal model capable of high-speed processing. Generates text from image, video, and text inputs. Can handle inputs up to 300K tokens and analyze multiple images and videos up to 30 minutes. Supports fine-tuning. |
| Amazon | Text & Vision | Nova Pro | A high-performance multimodal model with excellent balance of accuracy, speed, and cost. Supports inputs up to 300K tokens and achieves top-level performance in visual question answering and video understanding. Capable of executing complex workflows through API and tool integration. Supports fine-tuning. |
| Amazon | Image | Nova Canvas | A state-of-the-art model capable of high-quality image generation. Features precise control over style and content, inpainting (partial modification), outpainting (image extension), and background removal editing functions. Achieves high performance in evaluating image generation fidelity. |
| Amazon | Video | Nova Reel | A state-of-the-art model capable of generating professional-quality videos. Enables video generation from text or images, with control over visual style and pacing. Demonstrates excellent performance in video quality and consistency. |
Until now, Amazon as a model provider has offered the Amazon Titan series, but in addition to that, the Amazon Nova series has become available, providing text generation from multimodal inputs, image generation with editing functions, and video generation with control over visual style and pacing.
Also, Rerank models (Amazon Rerank, Cohere Rerank) to improve search accuracy in RAG applications have been provided and become available from Amazon and Cohere providers.
Serverless Models Announced at AWS re:Invent 2024 (Coming Soon)
Next, I'll introduce the Serverless models announced as coming soon at AWS re:Invent 2024.Among these, Stable Diffusion 3.5 Large became available on 2024-12-19.
| Provider | Modality | Model Name | Overview |
|---|---|---|---|
| Amazon | Text & Vision | Nova Premier | The top-tier multimodal model for complex inference tasks. Also optimal as a teacher model for knowledge distillation of custom models (knowledge transfer of probability distribution and latent representation of intermediate layers from large-scale models to small-scale models). Scheduled for release in early 2025. |
| poolside | Text | malibu | A model specialized for complex software engineering challenges such as advanced tasks including code generation, test creation, refactoring, and documentation creation. By collaborating with an assistant, it can be used directly within the developer's IDE, is fine-tuned based on the knowledge base, and has the flexibility to meet organization-specific needs. |
| poolside | Text | point | A model specialized in rapid code completion that accurately predicts developer needs using advanced context awareness. By collaborating with an assistant, it can be used directly within the developer's IDE, is fine-tuned based on the knowledge base, and has the flexibility to meet organization-specific needs. |
| Stability AI | Image | Stable Diffusion 3.5 Large | * Became available on 2024-12-19. The latest high-performance image generation model provided by Stability AI. Capable of generating high-quality and beautiful images from text. Streamlines the creation of concept art, visual effects, and detailed product images. |
| Luma AI | Video | Ray 2 | The latest video generation model that can generate high-quality videos from text or image prompts in about 10 seconds. Achieves smooth motion, advanced filming techniques, and dynamic camera work, capable of creating videos up to 1 minute long. |
Amazon Bedrock Marketplace Model List (as of 2024-12-16)
Finally, I'll introduce the models in the Amazon Bedrock Marketplace.As mentioned at the beginning, there were 122 types of models in the Amazon Bedrock Marketplace as of 2024-12-16, but the majority of these, 83 types, were models from HuggingFace, an open-source community.
In addition, various providers offer models with unique features, such as industry-specific models for medical and financial sectors.
| Provider | Category | Model Name |
|---|---|---|
| HuggingFace | Text Generation, Text Summarization, Automatic Speech Recognition, etc. | Number of models: 83 types. Main series: BART, Bloom, DBRX, Dolly, EleutherAI GPT, Falcon, Flan-T5, Gemma, Mistral, MPT, Phi, Yi, Zephyr, etc. |
| Arcee AI | Text Generation | Arcee[Lite, Nova, SuperNova], Llama Spark, Llama 3.1 SuperNova Lite |
| Camb.ai | Text To Audio | MARS6 |
| EvolutionaryScale, PBC | Multimodal Generation | ESM3-open |
| Gretel | Text Generation | Gretel Navigator Tabular |
| IBM Data and AI | Text Generation | IBM Granite[8B Code Instruct - 128K, 3B Code Instruct - 128K, 34B Code Instruct - 8K, 20B Code Instruct - 8K], Granite 3.0[8B Instruct, 2B Instruct] |
| John Snow Labs | Text Summarization | Medical LLM[Small, Medium] |
| John Snow Labs | Translation | Medical Text Translation (EN-ES) |
| Karakuri, Inc. | Text Generation | KARAKURI LM 8x7b instruct |
| LG CNS | Text Generation | EXAONE_v3.0 7.8B Instruct |
| Liquidai | Text Generation | Liquid LFM[40B (L40S), 40B (H100), 40B (A100)] |
| NCSoft | Text Generation | Llama-3-Varco-Offsetbias-8B, VARCO LLM KO/EN-13B-IST |
| NVIDIA | Text Generation | NVIDIA Nemotron-4 15B NIM Microservice |
| Preferred Networks, Inc. | Text Generation | PLaMo API |
| Stability AI | Text To Image | Stable Diffusion 3.5 Large |
| Stockmark Inc. | Text Generation | Stockmark-LLM-13b |
| Upstage | Text Generation | Solar[Pro, Pro – Quant], Solar Mini[Chat, Chat – Quant, Chat ja, Chat ja – Quant] |
| Widn.AI | Translation | Widn Tower Sugarloaf, Widn Tower Anthill, Widn Llama3-Tower Vesuvius |
| Writer | Text Generation | Writer Palmyra-Med-70B-32K, Writer Palmyra-Fin-70B-32K |
References:
Tech Blog with curated related content
Comprehensive Overview of Amazon Bedrock Models 2024(Japanese Presentation) | Speaker Deck
Comprehensive Overview of Amazon Bedrock Models 2024(Japanese Presentation) | Docswell
Summary
In this article, I summarized the overall framework of Amazon Bedrock announcements at AWS re:Invent 2024 and considered an overview of Amazon Bedrock models as of the end of 2024.For Amazon Bedrock as a whole at AWS re:Invent 2024, I felt that it strongly emphasized directions such as improving reliability and safety assuming enterprise use, enhancing performance and efficiency of the entire Bedrock utilization system, improving user and developer experience and lowering implementation barriers, and supporting a wider range of use cases.
Focusing on the Amazon Bedrock model updates at AWS re:Invent 2024, the major changes were the addition of the Amazon Nova series strengthening Amazon's proprietary AI models including multimodal capabilities, the addition of Rerank models (Amazon Rerank, Cohere Rerank) enabling improved RAG search accuracy, and the emergence of Amazon Bedrock Marketplace making even more diverse models available.
Also, examining Amazon Bedrock models as of the end of 2024 from various angles revealed that the cycle of feature additions and model additions/removals in Amazon Bedrock is rapid, and it's important to consider this in keeping up and building systems.
Especially for systems in the operational phase, I strongly felt the need to constantly try out and understand the latest models, assuming the replacement of currently used models.
I plan to continue keeping up with updates to Amazon Bedrock and AI models, and to try them out regularly to understand their functions, features, and use cases.
Written by Hidekazu Konishi
Copyright © Hidekazu Konishi ( hidekazu-konishi.com ) All Rights Reserved.