Amazon Bedrock Model Catalog 2026

First Published: 2026-05-16
Last Updated: 2026-05-16

* The contents of this article reflect the state as of the Last Updated date shown above. Bedrock catalog rows are revised in place as new models GA or are retired; check the model card before any production decision.

Snapshot scope: This is a point-in-time reference of every foundation model offered as a Bedrock-managed inference target.
Last reviewed: 2026-05.
Update cadence: refreshed as needed within the publication year; the URL stays stable across refreshes and each refresh overwrites this page in place.
Sources of truth: AWS Bedrock User Guide "Models at a glance", per-model cards, and the Regional Availability pages.

In late 2023 Amazon Bedrock launched with a single-digit number of model families. As of 2026, the catalog spans 18 providers and well over 110 individually addressable model variants, including 40+ language and reasoning models, three monolithic image generators plus 13 image-editing primitives from Stability AI, multiple speech and video models, and three inference routing tiers (In-Region, Geo, Global). Almost every model now ships with at least one cross-Region inference profile, so picking a model is no longer just "name plus Region" but "name plus Region plus inference profile prefix".

For engineers, architects, and PoC leads, the practical question is no longer "is the model on Bedrock?" but "which exact ID do I call, in which Region, with which capabilities switched on?" This page is built to answer that in one fetch.

1. How This Snapshot Was Compiled

1.1 Scope and Inclusion Criteria

This page covers every foundation model that is callable through the Amazon Bedrock bedrock-runtime inference endpoint as of the review date in the header. We include:

Generally Available (GA) and Active lifecycle models.
Models in Legacy lifecycle that still accept inference calls but have a published End-of-Life (EOL) date.
Preview models that are visible in the public model-cards index and accept inference (these are tagged [Preview]).

We exclude:

Models that exist only in the Bedrock Marketplace as third-party listings without a Bedrock-managed inference endpoint.
Models in Imported model form that customers bring themselves via Custom Model Import.
Anything restricted to non-public allowlists.

1.2 Sources of Truth

Every row in this catalog is cross-checked against three AWS sources:

The "Models at a glance" landing page (docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html).
The per-model "Model card" pages (one URL per model).
The "Regional availability" page (docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html) and the per-model Regional Availability tables.

Pricing is deliberately not included on this page. Bedrock token-rate pricing changes more frequently than this snapshot is regenerated, and stale numbers create more risk than value. For pricing, follow the Amazon Bedrock Pricing page.

1.3 Update Cadence

This page is refreshed as needed within the publication year. We retain the same URL on each refresh so that links from other pages, AI agents, and search engines do not rot. The Last reviewed date in the header is the single source of truth for "how fresh is this?".

1.4 Notation: Badges and Symbols

Throughout the tables we use the following compact notation:

[Preview] — the model is publicly visible but not yet GA. Quota and SLA may differ.
[Legacy] — the model is still callable but the provider has scheduled a successor; an EOL date is published.
[EOL YYYY-MM-DD] — the model will stop accepting inference on this date.
Yes / - — feature supported / not supported.
In-Region / Geo / Global — which inference routing tier(s) the model supports in a given Region (see §4.4).
n/a — the column does not apply (for example, Output length on an embedding model, or Context on a small on-demand LLM that is sized by request rather than by a published context cap).

A note on Model IDs: Bedrock uses several ID conventions in parallel and the differences are not cosmetic - the exact string is required by InvokeModel / Converse. Patterns to watch for:

Dated and revision-suffixed IDs: anthropic.claude-haiku-4-5-20251001-v1:0, mistral.pixtral-large-2502-v1:0, twelvelabs.pegasus-1-2-v1:0.
Clean suffix-less IDs for newer releases: anthropic.claude-opus-4-7, anthropic.claude-sonnet-4-6, mistral.devstral-2-123b, qwen.qwen3-vl-235b-a22b.
Dot-separated version numbers: zai.glm-4.7 (not zai.glm-4-7), minimax.minimax-m2.5, moonshotai.kimi-k2.5, deepseek.v3.2.
Double-namespace IDs: minimax.minimax-m2; same provider name appears in both the provider segment and the model name.
Two provider prefixes for one provider: Moonshot publishes Kimi K2 Thinking as moonshot.kimi-k2-thinking and Kimi K2.5 as moonshotai.kimi-k2.5. Always copy the exact prefix from the model card.
Implicit context-length variants: many models also expose ID aliases that pin the maximum context, e.g. amazon.nova-2-lite-v1:0:256k, amazon.nova-premier-v1:0:1000k, meta.llama3-3-70b-instruct-v1:0:128k. These behave like the base ID but with a hard context cap; the catalog tables list the unsuffixed base ID.

2. Quick Reference Summary

2.1 Model Count by Provider

The headline number for 2026 is that Bedrock now exposes models from 18 providers, a sharp jump from the 6 providers that defined the 2024 snapshot. Many of the additions are open-weight providers (Google Gemma, OpenAI's GPT-OSS family, Qwen, DeepSeek, Z.AI, Moonshot, MiniMax, NVIDIA Nemotron) whose models are hosted on Bedrock as Bedrock-managed inference rather than as fully proprietary APIs. Counts below include Active and Legacy GA models that still accept inference calls, but exclude older revision aliases such as ...-v1:0:200k.

* You can sort the table by clicking on the column name.

Provider	GA models in catalog
Stability AI	16
Amazon	16
Mistral AI	14
Meta	12
Anthropic	11
Qwen	7
Cohere	6
OpenAI	4
NVIDIA	4
Z.AI	3
MiniMax	3
Google	3
Writer	3
TwelveLabs	3
DeepSeek	3
Moonshot AI	2
AI21 Labs	2
Luma AI	1
Total	113

2.2 Model Count by Modality

* You can sort the table by clicking on the column name.

Modality	Count (approx.)	Examples
Text generation (chat / reasoning / coding)	70+	Claude Opus 4.7, Nova 2 Lite, Llama 4 Maverick, Mistral Large 3, Qwen3 235B A22B 2507, GLM 5
Vision (image input → text output)	20+	Claude 4.x family, Nova Pro / Premier / 2 Lite, Llama 3.2 Vision (Legacy) / Llama 4, Pixtral Large, Magistral Small 2509, Qwen3 VL, Gemma 3, Writer Palmyra Vision 7B
Image generation and editing	18	Stability AI suite (16: SD3.5 Large, Stable Image Core / Ultra, plus 13 editing primitives), Amazon Nova Canvas (Legacy), Titan Image Generator G1 v2 (Legacy)
Video understanding	4	Nova Premier, Nova 2 Lite, TwelveLabs Pegasus v1.2, TwelveLabs Marengo Embed (video input)
Video generation	3	Amazon Nova Reel (Legacy), Luma Ray v2, (Nova Reel earlier v1:0 build also still Legacy-listed)
Speech-to-speech (bidirectional)	2	Amazon Nova Sonic (Legacy), Amazon Nova 2 Sonic
Audio in → Text out	2	Mistral Voxtral Mini 3B 2507, Voxtral Small 24B 2507
Text embeddings	6	Cohere Embed (English v3 / Multilingual v3 / v4), Amazon Titan Text Embeddings V2, Amazon Titan Embeddings G1 - Text v1, Amazon Titan Embeddings G1 - Text v2 (`amazon.titan-embed-g1-text-02`)
Multimodal embeddings	4	Amazon Nova 2 Multimodal Embeddings, Amazon Titan Multimodal Embeddings G1, TwelveLabs Marengo Embed v2.7, TwelveLabs Marengo Embed 3.0
Rerank	2	Cohere Rerank 3.5, Amazon Rerank

2.3 Headline Numbers

Largest context window: 1,000,000 tokens (1M), shared by Anthropic Claude Opus 4.7, Amazon Nova Premier (Legacy), Amazon Nova 2 Lite, and Meta Llama 4 Maverick / Scout.
Largest output: 128K tokens, on Anthropic Claude Opus 4.7. The rest of the Claude 4.x family clusters at 64K output (32K on the original Claude Opus 4, now Legacy).
AWS Regions with at least one Bedrock model: 30+ commercial Regions when Global cross-Region inference is considered; far fewer (typically 4–8) when you constrain to In-Region inference for data residency.
Inference profile prefixes in use: us., eu., apac., jp., au., and global. (see §4.4).
Latest GA at review time: Claude Opus 4.7 (2026-04-16) and the Nova 2 generation (Nova 2 Lite, Nova 2 Sonic, Nova 2 Multimodal Embeddings).

3. Detailed Model Catalog

3.1 Bedrock Model Family Lineage

The catalog is wide enough that a single index helps orient the rest of this page. The table below groups models by provider and family axis (text vs. multimodal vs. image vs. speech vs. embedding/rerank).

* You can sort the table by clicking on the column name.

Provider	Family	Example models	Modality
Anthropic	Claude 4.x	Opus 4.7 / 4.6 / 4.5 / 4.1, Opus 4 [Legacy], Sonnet 4.6 / 4.5, Sonnet 4 [Legacy], Haiku 4.5	Text + Vision
Anthropic	Claude 3.x	3.5 Haiku [Legacy], 3 Haiku [Legacy]	Text + Vision
Amazon	Nova 2	Lite, Sonic, Multimodal Embeddings	Multimodal / Speech / Embedding
Amazon	Nova (Gen 1)	Premier [Legacy], Pro, Lite, Micro	Multimodal
Amazon	Nova Studio	Canvas [Legacy], Reel [Legacy], Sonic [Legacy]	Image / Video / Speech
Amazon	Titan	Image Gen G1 v2 [Legacy], Text Embeddings V2, MM Embeddings G1, Embed G1 Text v1 / v2, Rerank	Image / Embedding / Rerank
Meta	Llama 4	Maverick 17B, Scout 17B	Multimodal MoE
Meta	Llama 3.x	3.3 70B Instruct; 3.2 1B / 3B / 11B Vision / 90B Vision [all Legacy]; 3.1 8B / 70B, 3.1 405B [Legacy]; 3 8B / 70B Instruct	Text / Vision
Mistral AI	Mistral Large / Small / Mixtral / Pixtral	Large 3 (675B), Large 24.07, Large 24.02, Small, 7B, Mixtral 8x7B, Pixtral Large 25.02	Text / MoE / Multimodal
Mistral AI	Ministral 3 / Voxtral / Devstral / Magistral	Ministral 3 (3B / 8B / 14B), Voxtral Mini 3B / Small 24B, Devstral 2 123B, Magistral Small 2509	Text small / Audio / Code / Reasoning + Vision
Cohere	Command / Embed / Rerank	Command R+ [Legacy], Command R [Legacy], Embed English v3 / Multilingual v3 / v4, Rerank 3.5	Text / Embedding / Rerank
AI21 Labs	Jamba 1.5	Large, Mini	Hybrid SSM
Stability AI	Stable Image (monolithic + editing)	SD3.5 Large, Stable Image Core / Ultra, plus 13 editing primitives (Conservative / Creative / Fast Upscale, Inpaint, Outpaint, Erase Object, Remove Background, Control Sketch / Structure, Search Recolor / Replace, Style Guide / Transfer)	Image generation / editing
DeepSeek	V3 / R1	V3.2, V3, R1	Text / Reasoning
Google	Gemma 3	27B IT, 12B IT, 4B IT	Text + Vision (open weight)
MiniMax	M2	M2.5, M2.1, M2	Text + Multimodal
Moonshot AI	Kimi	K2.5, K2 Thinking	Text / Reasoning
NVIDIA	Nemotron	Super 3 120B A12B, Nano 3 30B, Nano 12B v2, Nano 9B v2	Text
OpenAI	GPT-OSS	gpt-oss-120b, gpt-oss-20b, Safeguard 120B, Safeguard 20B	Text (open weight) / Classification
Qwen	Qwen3	Qwen3 235B A22B 2507, Qwen3 VL 235B A22B, Qwen3 Next 80B A3B, Qwen3 32B, Coder Next / 480B A35B / 30B A3B	Text / Vision / Code (MoE)
Writer	Palmyra	Palmyra X5, X4, Vision 7B	Text / Vision
TwelveLabs	Marengo / Pegasus	Marengo Embed 3.0, Marengo Embed v2.7, Pegasus v1.2	Video understanding / embedding
Z.AI	GLM	GLM 5, GLM 4.7, GLM 4.7 Flash	Text
Luma AI	Ray	Ray v2	Video generation

3.2 Provider-Indexed Anchor List

Jump directly to a provider:

Anthropic - Claude Family
Amazon - Nova / Titan / Rerank
Meta - Llama Family
Mistral AI
Cohere
AI21 Labs - Jamba
Stability AI - Image Generation
DeepSeek
Google - Gemma
MiniMax
Moonshot AI - Kimi
NVIDIA - Nemotron
OpenAI - GPT-OSS
Qwen
Writer - Palmyra
TwelveLabs
Z.AI - GLM
Luma AI

3.3 Anthropic - Claude Family

The Claude family on Bedrock spans the GA Claude 4.x generation and a Claude 3 tail that has now largely transitioned to the Legacy lifecycle. Claude Opus 4.7 is the headline model at the review date, launched on 2026-04-16 with a 1M-token context window and 128K output. All Claude 4.x models support Tool Use, Prompt Caching, Guardrails, and Knowledge Bases.

Note on ID conventions: in 2026 Bedrock uses two ID forms for Claude in parallel. The newest releases (Opus 4.7, Opus 4.6, Sonnet 4.6) ship with cleaner suffix-less IDs; older releases keep dated, versioned IDs such as anthropic.claude-opus-4-5-20251101-v1:0. Both forms are accepted by InvokeModel / Converse.

* You can sort the table by clicking on the column name.

Model name	Model ID (`bedrock-runtime`)	Modality	Context	Output	Streaming	Tool Use	Prompt Caching	Cross-Region Inference	Lifecycle
Claude Opus 4.7	`anthropic.claude-opus-4-7`	Text + Vision	1M	128K	Yes	Yes	Yes (1K min, 4 ckpt, 5 min TTL)	us, eu, jp, au, global	Active
Claude Opus 4.6	`anthropic.claude-opus-4-6-v1`	Text + Vision	200K	64K	Yes	Yes	Yes	us, eu, au, global	Active
Claude Opus 4.5	`anthropic.claude-opus-4-5-20251101-v1:0`	Text + Vision	200K	64K	Yes	Yes	Yes	us, eu, global	Active
Claude Opus 4.1	`anthropic.claude-opus-4-1-20250805-v1:0`	Text + Vision	200K	64K	Yes	Yes	Yes	us	Active
Claude Opus 4	`anthropic.claude-opus-4-20250514-v1:0`	Text + Vision	200K	32K	Yes	Yes	Yes	us	Legacy
Claude Sonnet 4.6	`anthropic.claude-sonnet-4-6`	Text + Vision	200K	64K	Yes	Yes	Yes	us, eu, jp, au, global	Active
Claude Sonnet 4.5	`anthropic.claude-sonnet-4-5-20250929-v1:0`	Text + Vision	200K	64K	Yes	Yes	Yes	us, eu, jp, au, global	Active
Claude Sonnet 4	`anthropic.claude-sonnet-4-20250514-v1:0`	Text + Vision	200K	64K	Yes	Yes	Yes	us, eu, apac, global	Legacy
Claude Haiku 4.5	`anthropic.claude-haiku-4-5-20251001-v1:0`	Text + Vision	200K	64K	Yes	Yes	Yes (4K min, 4 ckpt, 5 min and 1 hour TTL)	us, eu, jp, global	Active
Claude 3.5 Haiku	`anthropic.claude-3-5-haiku-20241022-v1:0`	Text + Vision	200K	8K	Yes	Yes	Yes	us	Legacy
Claude 3 Haiku	`anthropic.claude-3-haiku-20240307-v1:0`	Text + Vision	200K	4K	Yes	Yes	-	us, eu, apac	Legacy

Note: Starting with Claude Opus 4.7, the temperature, top_p, and top_k sampling parameters are no longer supported — guide model behavior via prompting instead. Extended thinking on Opus 4.7 is thinking.type: "adaptive" only; the older thinking.type: "enabled" with budget_tokens returns a 400 error.

3.4 Amazon - Nova / Titan / Rerank

Amazon's own portfolio has the broadest modality span on Bedrock. Nova 2 (Lite, Sonic, Multimodal Embeddings) is the current generation, launched in late 2025 and early 2026. The first-generation Nova lineup (Premier, Pro, Lite, Micro, the original Sonic / Canvas / Reel) and the older Titan Image Generator have all transitioned to Legacy lifecycle as their Nova 2 successors GA. The Titan embedding models still anchor the long-tail of RAG workloads.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Output	Tool Use	Prompt Caching	Cross-Region Inference	Lifecycle
Nova 2 Lite	`amazon.nova-2-lite-v1:0`	Text + Vision + Video	1M (256K variant also published as `amazon.nova-2-lite-v1:0:256k`)	64K	Yes	Yes (1K min, 4 ckpt, 5 min)	us, eu, jp, global	Active
Nova 2 Sonic	`amazon.nova-2-sonic-v1:0`	Speech in → Speech + Text out (bidirectional)	n/a	n/a	n/a	-	In-Region (us-east-1, us-west-2, ap-northeast-1)	Active
Nova 2 Multimodal Embeddings	`amazon.nova-2-multimodal-embeddings-v1:0`	Embedding (text + image + video)	n/a	n/a	n/a	-	In-Region (us-east-1)	Active
Nova Premier	`amazon.nova-premier-v1:0`	Text + Vision + Video	1M	25K	Yes	Yes (1K min, 4 ckpt, 5 min)	us	Legacy (EOL 2026-09-14)
Nova Pro	`amazon.nova-pro-v1:0`	Text + Vision + Video	300K	5K	Yes	Yes	us, eu, apac	Active
Nova Lite	`amazon.nova-lite-v1:0`	Text + Vision + Video	300K	5K	Yes	Yes	us, eu, apac	Active
Nova Micro	`amazon.nova-micro-v1:0`	Text	128K	5K	Yes	Yes	us, eu, apac	Active
Nova Sonic	`amazon.nova-sonic-v1:0`	Speech in → Speech + Text out	n/a	n/a	n/a	-	In-Region (us-east-1, ap-northeast-1)	Legacy (succeeded by Nova 2 Sonic)
Nova Canvas	`amazon.nova-canvas-v1:0`	Image generation (text + image input)	n/a	n/a	n/a	-	In-Region (us-east-1, ap-northeast-1, eu-west-1)	Legacy (EOL 2026-09-30 in us-east-1)
Nova Reel	`amazon.nova-reel-v1:1`	Video generation (text + image input)	n/a	n/a	n/a	-	In-Region (us-east-1, ap-northeast-1, eu-west-1)	Legacy (a prior `nova-reel-v1:0` build is also still listed as Legacy)
Titan Image Generator G1 v2	`amazon.titan-image-generator-v2:0`	Image generation (text + image input)	n/a	n/a	n/a	-	In-Region (us-east-1, us-west-2)	Legacy
Titan Multimodal Embeddings G1	`amazon.titan-embed-image-v1`	Embedding (text + image)	n/a	n/a	n/a	-	In-Region (us-east-1, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-central-1, eu-west-1)	Active
Titan Text Embeddings V2	`amazon.titan-embed-text-v2:0`	Embedding (text)	8K input	n/a	n/a	-	In-Region in nearly every commercial Region	Active
Titan Embeddings G1 - Text v1	`amazon.titan-embed-text-v1`	Embedding (text)	8K input	n/a	n/a	-	In-Region (us-east-1, us-west-2, ap-northeast-1, eu-central-1, eu-west-1)	Active
Titan Embeddings G1 - Text v2 (alt build)	`amazon.titan-embed-g1-text-02`	Embedding (text)	8K input	n/a	n/a	-	In-Region (us-east-1, us-west-2, eu-central-1)	Active
Rerank	`amazon.rerank-v1:0`	Rerank	n/a	n/a	n/a	-	In-Region (us-west-2, ap-northeast-1, eu-central-1)	Active

3.5 Meta - Llama Family

Llama 4 introduced mixture-of-experts to the Bedrock-hosted Llama line. Maverick and Scout are both 17B-active MoEs with a 1M context window and multimodal (image) input. Llama 3.x covers a wider parameter range, including the dense 405B model in 3.1, though the entire 3.2 sub-family and Llama 3.1 405B have transitioned to Legacy as their successors stabilized.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Output	Tool Use	Cross-Region Inference	Lifecycle
Llama 4 Maverick 17B Instruct	`meta.llama4-maverick-17b-instruct-v1:0`	Text + Vision (MoE)	1M	8K	Yes	us	Active
Llama 4 Scout 17B Instruct	`meta.llama4-scout-17b-instruct-v1:0`	Text + Vision (MoE)	1M	8K	Yes	us	Active
Llama 3.3 70B Instruct	`meta.llama3-3-70b-instruct-v1:0`	Text	128K	8K	Yes	us	Active
Llama 3.2 90B Instruct (Vision)	`meta.llama3-2-90b-instruct-v1:0`	Text + Vision	128K	8K	Yes	us	Legacy
Llama 3.2 11B Instruct (Vision)	`meta.llama3-2-11b-instruct-v1:0`	Text + Vision	128K	8K	Yes	us	Legacy
Llama 3.2 3B Instruct	`meta.llama3-2-3b-instruct-v1:0`	Text	128K	8K	Yes	us, eu	Legacy
Llama 3.2 1B Instruct	`meta.llama3-2-1b-instruct-v1:0`	Text	128K	8K	Yes	us, eu	Legacy
Llama 3.1 405B Instruct	`meta.llama3-1-405b-instruct-v1:0`	Text	128K	8K	Yes	us	Legacy
Llama 3.1 70B Instruct	`meta.llama3-1-70b-instruct-v1:0`	Text	128K	8K	Yes	us	Active
Llama 3.1 8B Instruct	`meta.llama3-1-8b-instruct-v1:0`	Text	128K	8K	Yes	us	Active
Llama 3 70B Instruct	`meta.llama3-70b-instruct-v1:0`	Text	8K	8K	-	In-Region (us-east-1, us-west-2, ap-south-1)	Active
Llama 3 8B Instruct	`meta.llama3-8b-instruct-v1:0`	Text	8K	8K	-	In-Region (us-east-1, us-west-2, ap-south-1)	Active

3.6 Mistral AI

Mistral's Bedrock catalog now spans the dense Mistral Large 3 flagship (published with the explicit -675b-instruct suffix in the Bedrock model ID), a classic mixture-of-experts model (Mixtral 8x7B), the multimodal Pixtral Large, the Ministral 3 small-LLM line (3B / 8B / 14B), audio-capable Voxtral models, the Devstral coding model, and Magistral Small for reasoning. Magistral Small 2509 accepts image input in addition to text, and the Voxtral models accept speech alongside text input.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Tool Use	Cross-Region Inference	Lifecycle
Mistral Large 3	`mistral.mistral-large-3-675b-instruct`	Text	128K	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Mistral Large 24.07	`mistral.mistral-large-2407-v1:0`	Text	128K	Yes	In-Region (us-west-2)	Active
Mistral Large 24.02	`mistral.mistral-large-2402-v1:0`	Text	32K	Yes	In-Region (us-east-1, us-west-2, ap-southeast-2, ap-south-1, eu-west-1)	Active
Mistral Small	`mistral.mistral-small-2402-v1:0`	Text	32K	Yes	In-Region (us-east-1)	Active
Mistral 7B Instruct	`mistral.mistral-7b-instruct-v0:2`	Text	32K	-	In-Region (us-east-1, us-west-2, ap-southeast-2, ap-south-1, eu-west-1)	Active
Mixtral 8x7B Instruct	`mistral.mixtral-8x7b-instruct-v0:1`	Text (MoE)	32K	-	In-Region (us-east-1, us-west-2, ap-southeast-2, ap-south-1, eu-west-1)	Active
Pixtral Large 25.02	`mistral.pixtral-large-2502-v1:0`	Text + Vision	128K	Yes	us, eu	Active
Ministral 3 14B Instruct	`mistral.ministral-3-14b-instruct`	Text	n/a (small LLM, On-Demand)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Ministral 3 8B Instruct	`mistral.ministral-3-8b-instruct`	Text	n/a (small LLM, On-Demand)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Ministral 3 3B Instruct	`mistral.ministral-3-3b-instruct`	Text	n/a (small LLM, On-Demand)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Devstral 2 123B	`mistral.devstral-2-123b`	Text (coding)	n/a (On-Demand)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-central-1, eu-west-1)	Active
Magistral Small 2509	`mistral.magistral-small-2509`	Text + Vision (reasoning)	n/a (On-Demand)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Voxtral Mini 3B 2507	`mistral.voxtral-mini-3b-2507`	Speech + Text in → Text out	n/a (On-Demand)	-	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active
Voxtral Small 24B 2507	`mistral.voxtral-small-24b-2507`	Speech + Text in → Text out	n/a (On-Demand)	-	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-west-1)	Active

3.7 Cohere

The Cohere Command R / R+ generative pair has now transitioned to Legacy as customers migrate to Embed v4 plus stronger general-purpose models from other providers; the Embed and Rerank lines remain Active and form the backbone of many RAG stacks on Bedrock.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Tool Use	Cross-Region Inference	Lifecycle
Command R+	`cohere.command-r-plus-v1:0`	Text	128K	Yes	In-Region (us-east-1, us-west-2)	Legacy
Command R	`cohere.command-r-v1:0`	Text	128K	Yes	In-Region (us-east-1, us-west-2)	Legacy
Embed English v3	`cohere.embed-english-v3`	Embedding	n/a	n/a	In-Region (us-east-1, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-central-1, eu-west-1)	Active
Embed Multilingual v3	`cohere.embed-multilingual-v3`	Embedding	n/a	n/a	In-Region (us-east-1, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1, eu-central-1, eu-west-1)	Active
Embed v4	`cohere.embed-v4:0`	Embedding	n/a	n/a	us, eu, global (Geo + Global available)	Active
Rerank 3.5	`cohere.rerank-v3-5:0`	Rerank	n/a	n/a	In-Region (us-west-2, ap-northeast-1, eu-central-1)	Active

3.8 AI21 Labs - Jamba

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Tool Use	Cross-Region Inference	Lifecycle
Jamba 1.5 Large	`ai21.jamba-1-5-large-v1:0`	Text (Hybrid SSM/Transformer)	256K	Yes	In-Region (us-east-1)	Active
Jamba 1.5 Mini	`ai21.jamba-1-5-mini-v1:0`	Text (Hybrid SSM/Transformer)	256K	Yes	In-Region (us-east-1)	Active

3.9 Stability AI - Image Generation

Stability's 2026 catalog is built around two layers: a small set of monolithic text-to-image generators (Stable Diffusion 3.5 Large, Stable Image Core, Stable Image Ultra) plus a wide set of editing primitives where each model handles one operation (inpaint, outpaint, upscale, control, search-and-replace, style). Each model accepts text and/or image inputs and returns an image. The editing primitives are hosted in us-west-2 with a subset also in us-east-1; the three monolithic generators are us-west-2-only at the review date. US Geo cross-Region profiles (us.stability....) are available for the editing primitives.

Naming convention note: the Bedrock model IDs are mostly prefixed stability.stable-image-..., but five IDs are published without the image- infix: stability.stable-conservative-upscale-v1:0, stability.stable-creative-upscale-v1:0, stability.stable-fast-upscale-v1:0, stability.stable-outpaint-v1:0, and stability.stable-style-transfer-v1:0. Always copy the exact ID from the model card before wiring it into IaC.

* You can sort the table by clicking on the column name.

Model name	Model ID	Operation
Stable Diffusion 3.5 Large	`stability.sd3-5-large-v1:0`	Text-to-image (monolithic, high quality)
Stable Image Core	`stability.stable-image-core-v1:1`	Text-to-image (monolithic, fast)
Stable Image Ultra	`stability.stable-image-ultra-v1:1`	Text-to-image (monolithic, ultra quality)
Stable Image Conservative Upscale	`stability.stable-conservative-upscale-v1:0`	Upscale (faithful)
Stable Image Creative Upscale	`stability.stable-creative-upscale-v1:0`	Upscale (creative)
Stable Image Fast Upscale	`stability.stable-fast-upscale-v1:0`	Upscale (fast)
Stable Image Inpaint	`stability.stable-image-inpaint-v1:0`	Inpaint
Stable Image Outpaint	`stability.stable-outpaint-v1:0`	Outpaint
Stable Image Erase Object	`stability.stable-image-erase-object-v1:0`	Object erase
Stable Image Remove Background	`stability.stable-image-remove-background-v1:0`	Background remove
Stable Image Control Sketch	`stability.stable-image-control-sketch-v1:0`	Sketch-guided gen
Stable Image Control Structure	`stability.stable-image-control-structure-v1:0`	Structure-guided gen
Stable Image Search and Recolor	`stability.stable-image-search-recolor-v1:0`	Object recolor
Stable Image Search and Replace	`stability.stable-image-search-replace-v1:0`	Object replace
Stable Image Style Guide	`stability.stable-image-style-guide-v1:0`	Style transfer (guide)
Stable Image Style Transfer	`stability.stable-style-transfer-v1:0`	Style transfer (full)

3.10 DeepSeek

DeepSeek is on Bedrock as Bedrock-managed inference behind US-prefixed inference profiles. The Bedrock model IDs use dot-versioned naming (deepseek.v3.2) for the latest revision and the conventional -v1:0 suffix for the earlier ones.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Context	Tool Use	Cross-Region Inference	Lifecycle
DeepSeek V3.2	`deepseek.v3.2`	Text	128K (typical for V3 family)	Yes	In-Region (us-east-1, us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1)	Active
DeepSeek V3	`deepseek.v3-v1:0`	Text	128K	Yes	In-Region (us-east-2, us-west-2, ap-northeast-1, ap-southeast-2, ap-south-1)	Active
DeepSeek-R1	`deepseek.r1-v1:0`	Text (reasoning)	128K	Yes	us (Geo profile only; `us.deepseek.r1-v1:0`)	Active

3.11 Google - Gemma

Google Gemma 3 ships on Bedrock as open-weight chat-tuned models. All three sizes are instruction-tuned (IT); no pretrained-only (PT) build is exposed through Bedrock-managed inference at the review date. All three accept image input alongside text.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
Gemma 3 27B IT	`google.gemma-3-27b-it`	Text + Vision	Active
Gemma 3 12B IT	`google.gemma-3-12b-it`	Text + Vision	Active
Gemma 3 4B IT	`google.gemma-3-4b-it`	Text + Vision	Active

3.12 MiniMax

The MiniMax M2 family uses double-namespace IDs (minimax.minimax-...) on Bedrock; the version number is dot-separated rather than dash-separated. All three are Active.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
MiniMax M2.5	`minimax.minimax-m2.5`	Text + Multimodal	Active
MiniMax M2.1	`minimax.minimax-m2.1`	Text + Multimodal	Active
MiniMax M2	`minimax.minimax-m2`	Text + Multimodal	Active

3.13 Moonshot AI - Kimi

Moonshot publishes Kimi K2 Thinking under the moonshot. provider namespace, while Kimi K2.5 was added later under the moonshotai. namespace and with dot-separated versioning. Both forms are valid Bedrock IDs - this is a deliberate inconsistency you must respect verbatim.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
Kimi K2.5	`moonshotai.kimi-k2.5`	Text	Active
Kimi K2 Thinking	`moonshot.kimi-k2-thinking`	Text (reasoning)	Active

3.14 NVIDIA - Nemotron

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
NVIDIA Nemotron 3 Super 120B A12B	`nvidia.nemotron-super-3-120b`	Text	Active
Nemotron Nano 3 30B	`nvidia.nemotron-nano-3-30b`	Text	Active
Nemotron Nano 12B v2	`nvidia.nemotron-nano-12b-v2`	Text	Active
NVIDIA Nemotron Nano 9B v2	`nvidia.nemotron-nano-9b-v2`	Text	Active

3.15 OpenAI - GPT-OSS

OpenAI's open-weight GPT-OSS models arrived on Bedrock in 2025, hosted as Bedrock-managed inference. The Safeguard variants are tuned for content classification rather than generation. The base GPT-OSS IDs carry an explicit -1:0 revision suffix; the Safeguard IDs do not.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
gpt-oss-120b	`openai.gpt-oss-120b-1:0`	Text (open weight)	Active
gpt-oss-20b	`openai.gpt-oss-20b-1:0`	Text (open weight)	Active
GPT OSS Safeguard 120B	`openai.gpt-oss-safeguard-120b`	Text (classification)	Active
GPT OSS Safeguard 20B	`openai.gpt-oss-safeguard-20b`	Text (classification)	Active

3.16 Qwen

Qwen3 on Bedrock uses two ID conventions in parallel: the larger newest releases (qwen3-vl-235b-a22b, qwen3-next-80b-a3b, qwen3-coder-next) ship without a -v1:0 suffix, while the dated and Instruct variants carry the conventional -v1:0.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
Qwen3 235B A22B 2507	`qwen.qwen3-235b-a22b-2507-v1:0`	Text (MoE)	Active
Qwen3 VL 235B A22B	`qwen.qwen3-vl-235b-a22b`	Text + Vision (MoE)	Active
Qwen3 Next 80B A3B	`qwen.qwen3-next-80b-a3b`	Text (MoE)	Active
Qwen3 Coder Next	`qwen.qwen3-coder-next`	Text (coding)	Active
Qwen3 Coder 480B A35B	`qwen.qwen3-coder-480b-a35b-v1:0`	Text (coding, MoE)	Active
Qwen3 Coder 30B A3B	`qwen.qwen3-coder-30b-a3b-v1:0`	Text (coding)	Active
Qwen3 32B	`qwen.qwen3-32b-v1:0`	Text	Active

3.17 Writer - Palmyra

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
Palmyra X5	`writer.palmyra-x5-v1:0`	Text	Active
Palmyra X4	`writer.palmyra-x4-v1:0`	Text	Active
Palmyra Vision 7B	`writer.palmyra-vision-7b`	Text + Vision	Active

3.18 TwelveLabs

Marengo Embed accepts all four modalities (text, image, speech, video) and returns an embedding vector. Pegasus v1.2 takes text + video and returns text (video understanding). All three are accessed through inference profiles (us., eu., global.).

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
Marengo Embed 3.0	`twelvelabs.marengo-embed-3-0-v1:0`	Text + Image + Speech + Video → Embedding	Active
Marengo Embed v2.7	`twelvelabs.marengo-embed-2-7-v1:0`	Text + Image + Speech + Video → Embedding	Active
Pegasus v1.2	`twelvelabs.pegasus-1-2-v1:0`	Text + Video → Text (video understanding)	Active

3.19 Z.AI - GLM

Z.AI publishes GLM with dot-separated versioning (zai.glm-4.7, not zai.glm-4-7). GLM 4.7 Flash is the only Z.AI model available in eu-central-1 at the review date.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Lifecycle
GLM 5	`zai.glm-5`	Text	Active
GLM 4.7	`zai.glm-4.7`	Text	Active
GLM 4.7 Flash	`zai.glm-4.7-flash`	Text (fast)	Active

3.20 Luma AI

Luma joined Bedrock in 2026 with a single video-generation model, currently us-west-2-only.

* You can sort the table by clicking on the column name.

Model name	Model ID	Modality	Cross-Region Inference	Lifecycle
Luma Ray v2	`luma.ray-v2:0`	Video generation (text + image input)	In-Region (us-west-2)	Active

4. Region Availability Matrix

Bedrock now offers three inference routing tiers. In-Region keeps the request inside a single Region (strict data residency). Geo routes across Regions within a geography (US, EU, APAC, JP, AU). Global routes anywhere worldwide. Most modern models are available in at least one Geo profile and a Global profile; older models are In-Region only.

The matrix below covers 8 representative Regions out of the 30+ Regions where Bedrock has presence. For the complete list, follow the per-model card link.

Legend: In = In-Region available; Geo = available via a Geo cross-Region profile sourced from this Region; Glb = available via the Global profile sourced from this Region; - = not available.

4.1 Text and Vision Models × 8 Regions

* You can sort the table by clicking on the column name.

Model	us-east-1	us-east-2	us-west-2	ap-northeast-1	ap-southeast-2	ap-south-1	eu-central-1	eu-west-1
Claude Opus 4.7	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Glb	In + Glb	In + Geo + Glb	In + Geo + Glb
Claude Haiku 4.5	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Glb	In + Glb	In + Geo + Glb	In + Geo + Glb
Claude Sonnet 4.6	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Glb	In + Geo + Glb	In + Geo + Glb
Claude Sonnet 4.5	In + Geo + Glb	Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Glb	In + Geo + Glb	In + Geo + Glb
Claude Sonnet 4 [Legacy]	In + Geo + Glb	Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb
Nova 2 Lite	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Geo + Glb	In + Glb	In + Glb	In + Geo + Glb	In + Geo + Glb
Nova Pro	In + Geo	In + Geo	-	In + Geo	In + Geo	In + Geo	-	In
Nova Lite	In + Geo	In + Geo	-	In + Geo	In + Geo	In + Geo	-	In
Nova Micro	In + Geo	In + Geo	-	-	-	-	-	In
Llama 4 Maverick 17B	In + Geo (us)	In + Geo (us)	In + Geo (us)	-	-	-	-	-
Llama 4 Scout 17B	In + Geo (us)	In + Geo (us)	In + Geo (us)	-	-	-	-	-
Llama 3.3 70B Instruct	In + Geo (us)	In + Geo (us)	In + Geo (us)	-	-	-	-	-
Mistral Large 3 (675B)	In	In	In	In	In	In	-	In
Mistral Large 24.07	-	-	In	-	-	-	-	-
Pixtral Large 25.02	In + Geo (us)	-	In + Geo (us)	-	-	-	In + Geo (eu)	In + Geo (eu)
Command R+ [Legacy]	In	-	In	-	-	-	-	-
Jamba 1.5 Large	In	-	-	-	-	-	-	-
DeepSeek V3.2	In	In	In	In	In	In	-	-
DeepSeek-R1	In + Geo (us)	In + Geo (us)	In + Geo (us)	-	-	-	-	-

4.2 Image, Video, and Speech Models × 8 Regions

* You can sort the table by clicking on the column name.

Model	us-east-1	us-east-2	us-west-2	ap-northeast-1	ap-southeast-2	ap-south-1	eu-central-1	eu-west-1
Nova Canvas [Legacy]	In `[EOL 2026-09-30]`	-	-	In	-	-	-	In
Nova Reel [Legacy]	In (v1:1 and v1:0)	-	-	In (v1:0)	-	-	-	In (v1:0)
Nova Sonic [Legacy]	In	-	-	In	-	-	-	-
Nova 2 Sonic	In	-	In	In	-	-	-	-
Titan Image Generator G1 v2 [Legacy]	In	-	In	-	-	-	-	-
Stable Image suite *	In (editing subset)	-	In (full; incl. SD3.5 Large / Core / Ultra)	-	-	-	-	-
Luma Ray v2	-	-	In	-	-	-	-	-

* The Stable Image suite is a family of 16 models on Bedrock: three monolithic text-to-image generators (SD3.5 Large, Stable Image Core, Stable Image Ultra) plus 13 per-operation editing primitives (upscale, inpaint, outpaint, erase, remove background, control, search-and-replace, style). The full 16 are hosted in us-west-2; the editing primitives also have an In-Region presence in us-east-1 and corresponding us.stability.... Geo cross-Region profiles. Per-primitive availability varies - confirm against the Bedrock Console Foundation models → Model access view before designing a workload.

4.3 Embedding and Rerank Models × 8 Regions

* You can sort the table by clicking on the column name.

Model	us-east-1	us-east-2	us-west-2	ap-northeast-1	ap-southeast-2	ap-south-1	eu-central-1	eu-west-1
Titan Text Embeddings V2	In	In	In	In	In	In	In	-
Titan Embeddings G1 - Text	In	-	-	In	-	-	In	In
Titan Multimodal Embeddings G1	-	-	-	-	In	In	In	In
Nova Multimodal Embeddings	In	-	-	-	-	-	-	-
Cohere Embed English	In	-	In	In	-	-	In	-
Cohere Embed Multilingual	In	-	In	In	-	-	In	-
Cohere Rerank 3.5	-	-	In	In	-	-	In	-
Amazon Rerank	-	-	In	In	-	-	In	-

4.4 Cross-Region Inference Profiles

Cross-Region inference profile IDs use a stable prefix that encodes the destination geography. As of 2026 the following prefixes are in active use:

Prefix	Geography	Typical destination Regions
`us.<model-id>`	US	us-east-1, us-east-2, us-west-2 (and us-west-1 / ca-central-1 as routed sources)
`eu.<model-id>`	Europe	eu-central-1, eu-north-1, eu-south-1, eu-south-2, eu-west-1, eu-west-3
`apac.<model-id>`	Asia Pacific (broad)	ap-northeast-1, ap-southeast-1, ap-southeast-2, ap-south-1 (and others)
`jp.<model-id>`	Japan	ap-northeast-1 (Tokyo), ap-northeast-3 (Osaka)
`au.<model-id>`	Australia / NZ	ap-southeast-2 (Sydney), ap-southeast-4 (Melbourne), ap-southeast-6 (New Zealand)
`global.<model-id>`	Worldwide	All commercial Regions; routes are model-dependent

Two important nuances:

The destination set of a Global profile can change over time as AWS adds Regions. The destination set of a Geo profile (us., eu., jp., etc.) is fixed.
Some inference profiles route differently depending on the source Region. For example, calling us.anthropic.claude-3-haiku-20240307-v1:0 from us-east-2 can route to us-east-1, us-east-2, or us-west-2; calling it from us-west-2 can only route to us-east-1 or us-west-2. Use GetInferenceProfile from the source Region to confirm.

A representative subset of profile IDs (verified against list-inference-profiles at the review date):

Model	Geo profile ID(s)	Global profile ID
Claude Opus 4.7	`us.anthropic.claude-opus-4-7`, `eu.anthropic.claude-opus-4-7`, `jp.anthropic.claude-opus-4-7`	`global.anthropic.claude-opus-4-7`
Claude Opus 4.6	`us.anthropic.claude-opus-4-6-v1`, `eu.anthropic.claude-opus-4-6-v1`, `au.anthropic.claude-opus-4-6-v1`	`global.anthropic.claude-opus-4-6-v1`
Claude Haiku 4.5	`us.anthropic.claude-haiku-4-5-20251001-v1:0`, `eu.anthropic.claude-haiku-4-5-20251001-v1:0`, `jp.anthropic.claude-haiku-4-5-20251001-v1:0`	`global.anthropic.claude-haiku-4-5-20251001-v1:0`
Claude Sonnet 4.6	`us.anthropic.claude-sonnet-4-6`, `eu.anthropic.claude-sonnet-4-6`, `jp.anthropic.claude-sonnet-4-6`, `au.anthropic.claude-sonnet-4-6`	`global.anthropic.claude-sonnet-4-6`
Claude Sonnet 4 [Legacy]	`us.anthropic.claude-sonnet-4-20250514-v1:0`, `eu.anthropic.claude-sonnet-4-20250514-v1:0`, `apac.anthropic.claude-sonnet-4-20250514-v1:0`	`global.anthropic.claude-sonnet-4-20250514-v1:0`
Nova Premier [Legacy]	`us.amazon.nova-premier-v1:0`	not supported
Nova 2 Lite	`us.amazon.nova-2-lite-v1:0`, `eu.amazon.nova-2-lite-v1:0`, `jp.amazon.nova-2-lite-v1:0`	`global.amazon.nova-2-lite-v1:0`
Nova Pro / Lite / Micro	`us.amazon.nova-{pro,lite,micro}-v1:0`, `eu.amazon.nova-{pro,lite,micro}-v1:0`, `apac.amazon.nova-{pro,lite,micro}-v1:0`	not supported
Llama 4 Maverick / Scout 17B	`us.meta.llama4-{maverick,scout}-17b-instruct-v1:0`	not supported
DeepSeek-R1	`us.deepseek.r1-v1:0`	not supported
Cohere Embed v4	`us.cohere.embed-v4:0`, `eu.cohere.embed-v4:0`	`global.cohere.embed-v4:0`
TwelveLabs Marengo Embed 3.0	`us.twelvelabs.marengo-embed-3-0-v1:0`, `eu.twelvelabs.marengo-embed-3-0-v1:0`	not supported
TwelveLabs Pegasus v1.2	`us.twelvelabs.pegasus-1-2-v1:0`, `eu.twelvelabs.pegasus-1-2-v1:0`	`global.twelvelabs.pegasus-1-2-v1:0`
Mistral Pixtral Large	`us.mistral.pixtral-large-2502-v1:0`, `eu.mistral.pixtral-large-2502-v1:0`	not supported

Notable absences: as of the review date, Claude Sonnet 4.6 does not ship an apac. profile (only jp. within APAC). The apac. profile family for Anthropic is concentrated on the Claude 3.x line and Claude Sonnet 4 (Legacy); the Sonnet 4.6 / 4.5 generation jumped directly to the more granular jp. / au. profiles.

5. Recent Additions (2026-02 to 2026-05)

The following models were added or promoted to GA on Bedrock in the rolling three-month window preceding the review date:

Anthropic Claude Opus 4.7 — GA 2026-04-16. First Bedrock Claude with a 1M context window and 128K max output. Adaptive-only extended thinking; sampling parameters dropped.
Amazon Nova 2 Lite — GA 2025-12-02 (carried over into early 2026). The first member of the Nova 2 generation, with multimodal in (text + image + video), 1M context, 64K output, In-Region in all 8 representative Regions, plus us., eu., jp., and Global cross-Region profiles.
Amazon Nova 2 Sonic — GA early 2026. Bidirectional speech-in / speech+text-out model; replaces Nova Sonic (now Legacy) for new builds.
Amazon Nova 2 Multimodal Embeddings (amazon.nova-2-multimodal-embeddings-v1:0) — new in early 2026, supersedes the original Nova Multimodal Embeddings ID. Currently In-Region (us-east-1) only.
Luma AI Ray v2 (luma.ray-v2:0) — brings Luma as the 18th Bedrock provider. Video generation, us-west-2-only at the review date.
MiniMax M2.5 / Moonshot Kimi K2.5 / Z.AI GLM 5 / Qwen3 expansion / NVIDIA Nemotron 3 (Super 120B A12B + Nano 3 30B / 9B v2 / 12B v2) — first-half 2026 expansion of open-weight providers on Bedrock-managed inference.
Stability AI — the legacy single Stable Diffusion XL model has been replaced by a 16-model lineup: three monolithic generators (Stable Diffusion 3.5 Large, Stable Image Core, Stable Image Ultra) plus 13 per-operation editing primitives.
Mistral Large 3 (mistral.mistral-large-3-675b-instruct) — new 675B Mistral flagship; In-Region in 7 of the 8 representative Regions at the review date.

6. Recent Deprecations and Scheduled EOLs

The following models are on a published EOL schedule or have already transitioned to Legacy. If you are calling any of these, plan a migration before the date shown (or before the next refresh in cases where no firm EOL is published yet):

Amazon Nova Premier — Lifecycle: Legacy. EOL: 2026-09-14. Successor: Nova 2 Lite (and a future Nova 2 Pro / Premier when GA). Same 1M context, broader Region availability, and Global routing.
Amazon Nova Canvas in us-east-1 — Legacy, EOL: 2026-09-30. The Tokyo (ap-northeast-1) and Ireland (eu-west-1) deployments are also Legacy. Plan to migrate to a successor image model (Stability AI Stable Image Ultra / Core / SD3.5 Large are the natural Bedrock-native successors).
Amazon Nova Reel (both v1:0 and v1:1 builds) — Legacy across all Regions; nova-reel-v1:0 in us-east-1 has the firm EOL 2026-09-30. Successor for video generation: Luma Ray v2 (luma.ray-v2:0) in us-west-2.
Amazon Nova Sonic (Gen 1) — Legacy; successor is Nova 2 Sonic.
Amazon Titan Image Generator G1 v2 — Legacy; same successor guidance as Nova Canvas.
Anthropic Claude 3.5 Haiku and Claude 3 Haiku — both transitioned to Legacy. The natural successor is Claude Haiku 4.5.
Anthropic Claude Opus 4 (20250514) and Claude Sonnet 4 (20250514) — both transitioned to Legacy as the Opus 4.5 / 4.6 / 4.7 and Sonnet 4.5 / 4.6 lineage stabilized.
Cohere Command R / Command R+ — both transitioned to Legacy; Cohere's Active surface on Bedrock is now Embed (English v3 / Multilingual v3 / v4) and Rerank 3.5.
Meta Llama 3.2 family (1B / 3B / 11B / 90B Vision) and Llama 3.1 405B — all Legacy. Active Meta successors are Llama 4 Maverick / Scout and Llama 3.3 70B Instruct.

Older Bedrock models that have already been removed from the catalog (and are therefore not listed in §3) include the original Claude 2 / 2.1, Claude Instant, Claude 3 Opus, Llama 2 family, Mistral Large v1, the original Stable Diffusion XL v1, and Titan Text G1 - Express / Lite. Note that some legacy Claude 3.x Sonnet IDs (claude-3-sonnet-20240229, claude-3-5-sonnet-20240620, claude-3-5-sonnet-20241022-v2:0, claude-3-7-sonnet-20250219-v1:0) still appear in regional list-foundation-models output for ap-south-1 in Legacy status; only the most current Claude line (Haiku 4.5 + Sonnet 4.x + Opus 4.x) is included in §3.3.

7. Frequently Asked Questions

7.1 What is the largest context window available on Bedrock as of 2026?

1,000,000 tokens (1M). Five models share this: Anthropic Claude Opus 4.7, Amazon Nova Premier (Legacy), Amazon Nova 2 Lite, Meta Llama 4 Maverick, and Meta Llama 4 Scout. For maximum output, Claude Opus 4.7 leads at 128K tokens; the rest of the Active Claude 4.x family clusters at 64K output, with the Legacy Claude Opus 4 (20250514) capped at 32K.

7.2 Which models support Tool Use / Function Calling as of 2026?

Tool Use is supported across essentially the entire 2025-2026 wave of GA models: the full Claude 3.x and 4.x line, all Nova text variants (Premier, Pro, Lite, Micro, Nova 2 Lite), Llama 3.1 and later (including Llama 4 Maverick / Scout), Mistral Large family (Large 3 / 24.07 / 24.02), Pixtral Large, Ministral 3 (3B / 8B / 14B), Devstral 2 123B, Magistral Small 2509, AI21 Jamba 1.5 Large / Mini, DeepSeek-R1 / V3 / V3.2, and most open-weight additions from 2026 (Qwen3, GLM, Kimi K2 Thinking, Nemotron 3 / Nano line). Cohere Command R / R+ also still advertise Tool Use but are now Legacy, so plan a migration. Tool Use is not supported on the older Llama 3, Mistral 7B / Mixtral 8x7B, embedding models, image generation models, speech models, or the Voxtral audio models.

7.3 Which models support Prompt Caching as of 2026?

Prompt caching support is concentrated in the Anthropic and Amazon Nova families:

Claude Opus 4.7, Opus 4.6, Opus 4.5, Opus 4.1, Opus 4 (Legacy), Sonnet 4.6, Sonnet 4.5, Sonnet 4 (Legacy), Haiku 4.5 — all support prompt caching. Claude Haiku 4.5 in particular supports both a 5-minute and a 1-hour TTL.
Claude 3.5 Haiku (now Legacy) also supports prompt caching.
Nova Premier (Legacy), Nova Pro, Nova Lite, Nova Micro, and Nova 2 Lite all support prompt caching. For Nova Premier and Nova 2 Lite the published limits are 1K minimum tokens per checkpoint, up to 4 checkpoints, and a 5-minute TTL (primarily for text prompts); the Pro / Lite / Micro variants are listed as supported without per-checkpoint figures — confirm against the per-model card before sizing a workload.

Other model families generally do not (yet) advertise prompt caching on Bedrock.

7.4 Which models are only available via Cross-Region Inference?

In 2026 the picture is more nuanced than "Geo-only" - most flagship models advertise both In-Region and one or more Geo / Global profiles, and the API call site is what determines whether the request stays inside one Region or routes across a geography. Models that only route via Cross-Region inference profiles at the review date:

Anthropic Claude Opus 4.1 (20250805-v1:0) — us. Geo profile only; no In-Region invocation.
Amazon Nova Premier [Legacy] — us.amazon.nova-premier-v1:0 only; no Global profile, no In-Region.
Meta Llama 4 Maverick / Scout, Llama 3.3 70B, Llama 3.x Instruct family — us. Geo profile only (the model also lists In-Region in the same source Regions for the profile-aware invocation path).
DeepSeek-R1 — us.deepseek.r1-v1:0 only; no Global profile.
TwelveLabs Marengo Embed 3.0 / Pegasus v1.2 — Geo profiles (us., eu.) only; Pegasus also has a Global profile.

Models that do support direct In-Region invocation in their flagship Regions include all of the Anthropic Claude Opus 4.5 / 4.6 / 4.7 line, Claude Sonnet 4.5 / 4.6, Claude Haiku 4.5, Nova 2 Lite, Cohere Embed v4, Pixtral Large, Mistral Large 3, Mistral Ministral 3 / Devstral / Magistral / Voxtral, NVIDIA Nemotron, OpenAI GPT-OSS, Qwen3, Z.AI GLM, MiniMax M2 family, Moonshot Kimi, and the Stability AI family.

In contrast, almost all embedding, rerank, image generation, video generation, and speech models are In-Region only and do not offer cross-Region routing. Exceptions to this rule are TwelveLabs Marengo / Pegasus (Geo / Global), Cohere Embed v4 (Geo / Global), and Stability AI's editing primitives (us. Geo only).

7.5 Which models have been deprecated in the most recent refresh?

The notable shifts in the most recent refresh:

Anthropic Claude 3.5 Haiku and Claude 3 Haiku moved to Legacy across all Regions where they were previously Active.
Anthropic Claude Opus 4 and Claude Sonnet 4 (both 20250514) moved to Legacy as Opus 4.5 / 4.6 / 4.7 and Sonnet 4.5 / 4.6 became the primary release line.
Cohere Command R and Command R+ moved to Legacy; Cohere's Active surface is now Embed (v3 English / v3 Multilingual / v4) and Rerank 3.5.
Amazon Nova Sonic (Gen 1) moved to Legacy in favor of Nova 2 Sonic.
Amazon Nova Canvas / Nova Reel / Titan Image Generator G1 v2 moved to Legacy across all Regions where they exist.
Amazon Nova Premier moved to Legacy with a 2026-09-14 EOL.
Meta Llama 3.2 (1B / 3B / 11B Vision / 90B Vision) and Llama 3.1 405B moved to Legacy as Llama 4 Maverick / Scout and Llama 3.3 70B took over the Active surface.

If you are calling any of these directly in production, plan migration on the timeline above. The successor for Nova Premier is Nova 2 Lite (and the upcoming Nova 2 Pro / Premier when they GA). The successor for the Nova Canvas / Reel image and video models is Stability AI's image suite plus Luma Ray v2 for video. The successor for the Claude 3 / 3.5 Haiku tail is Claude Haiku 4.5.

8. References

8.1 AWS Documentation (External)

8.2 Internal Articles

8.3 Update Policy

This snapshot is refreshed as needed within the publication year. Last reviewed: 2026-05. The URL is intentionally stable across refreshes — link to it from your own documentation and the link will continue to resolve.

References:
Tech Blog with curated related content

Written by Hidekazu Konishi