Bedrock Llama end of life 2027: 26 open-weight models retire 30 March

Summary. Amazon Bedrock moved 26 open-weight models from nine providers into the Legacy state on 30 September 2026 and ends them on 30 March 2027, 143 days after 7 November 2026. The list covers every remaining Llama model in Bedrock's commercial Regions (Llama 3, 3.1, 3.3 and 4), plus DeepSeek, Gemma 3, Mistral AI, Qwen3 Coder, Kimi, Nemotron Nano, GLM and MiniMax models; 20 are listed for Mumbai (ap-south-1). Legacy already applies: customers new to a model cannot start, existing ones can lose access after 15 idle days, and no new Provisioned Throughput or fine-tuning is allowed. The higher-priced public extended access phase is undated and cannot start before 30 December 2026. AWS names no replacements. A Mumbai Llama 3 70B workload pays $3.18 input and $4.20 output per million tokens; in-Region alternatives include Mistral Large 3 ($0.59/$1.76), MiniMax M2.5 ($0.36/$1.44), Nemotron 3 Super 120B ($0.18/$0.78) and gpt-oss-120b ($0.18/$0.71). Custom Model Import keeps Llama 3.x weights, but only in US Regions and Frankfurt. Start testing in early January and cut over by mid-March.

About this article. This guide is for platform engineers, ML leads and CTOs who run open-weight models on Amazon Bedrock. We checked every model name, ID, Region, date, limit and price against AWS's own pages on 7 November 2026: the Bedrock Legacy lifecycle page as published that day, the model cards, the pricing page and the price data file behind it, the Custom Model Import guide and the API reference. AWS revises these pages often, so re-check before you commit. This is general information, not legal or financial advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What AWS changed on 30 September 2026, and the dates that matter

The Bedrock Legacy lifecycle page governs every model launched before 7 September 2026, and all 26 share three dates.

Four Legacy rules change how you run these models now, not only in March:

• New customers are blocked. AWS does not say whether a new account in an existing customer's organisation counts as existing, so do not assume it does.

• Fifteen idle days can end access early. Monthly batch jobs and idle staging accounts are exposed.

• No new Provisioned Throughput or fine-tuning jobs. Deployments created before Legacy keep working, and a model customised earlier can still get a new on-demand deployment.

• Nothing migrates for you. After end of life, requests fail unless you have a private arrangement with the provider.

Public extended access starts after at least 3 months in Legacy, and AWS promises at least 3 months' notice of the date. The first rule sets the floor at 30 December 2026; read literally, the second means a date announced after 7 November could not start before February 2027. Claude Sonnet 4.5, Legacy since 8 October 2026, follows the pattern with extended access from 8 January 2027, covered in our Bedrock Claude Opus 4.1 and Sonnet 4.5 retirement guide.

Lifecycle state is set per Region, and a model can stay Active in Regions its retirement entry does not list, so read the modelLifecycle field in each Region you use. For how this compares with OpenAI, Anthropic, Google and Microsoft, see our AI model retirement policies comparison.

The full inventory: all 26 models, as AWS lists them

Names and IDs come from the AWS Legacy lifecycle page, including second IDs for the OpenAI-compatible bedrock-mantle endpoint; the last column shows whether the retirement entry lists Mumbai.

Two details are easy to miss. AWS names one model Gemma 3 27B PT, but its ID ends in -it, so search code for the ID, not the label. And this batch empties Meta's row: Llama 3.1 405B and the four Llama 3.2 models reached end of life on 7 July 2026, and every other Llama on the models at a glance page is in this batch. Apart from AWS GovCloud (US-West), which the Llama 3 70B card lists but the retirement entry does not cover, no Meta model remains on managed Bedrock after 30 March unless AWS adds one.

Find out which of the 26 you actually call

Teams know their main model ID; fewer know about the eval harness, flow, prompt template, agent or nightly summariser that also calls a retiring model. Check four sources:

1. Code and configuration. Search repositories, infrastructure-as-code and parameter stores for each ID, including the mantle variants.

1. CloudWatch. Bedrock invocation metrics by ModelId over 90 days show what is called, and how often.

1. CloudTrail. Runtime call records identify the IAM principal, which maps calls to teams.

1. The lifecycle API. ListFoundationModels returns each model's lifecycle status and end-of-life time; the script below lists Legacy models per Region to diff against the first three.

            import boto3

REGIONS = ["ap-south-1", "us-east-1", "us-east-2", "us-west-2", "eu-west-2"]

for region in REGIONS:
    bedrock = boto3.client("bedrock", region_name=region)
    for m in bedrock.list_foundation_models()["modelSummaries"]:
        life = m.get("modelLifecycle", {})
        if life.get("status") == "LEGACY":
            print(region, m["providerName"], m["modelId"], life.get("endOfLifeTime"))
          

Flag any caller that runs less often than every 15 days. Either schedule a light keep-alive call until cutover or migrate it first.

Replacement candidates on Bedrock

AWS names no replacement for any of the 26. The candidates below are our editorial shortlist, chosen by family, modality and Region, and are starting points for your own evaluation.

The catch: replacements have their own clocks

Candidates launched before 7 September 2026 sit on the same legacy policy, and each card shows an earliest end-of-life date. gpt-oss-120b passed its date on 5 August 2026; DeepSeek V3.2 and Mistral Large 3 pass theirs on 1 and 2 December 2026. After that, only the 6-month Legacy notice protects you, so a model moved to Legacy in December would still run until at least June 2027. Newer models follow the card-based policy, and GLM 5.3 and Kimi K3 promise only 45 days.

Dates and Regions come from each model card, and Standard-tier prices from the pricing page and its price data.

AWS prices the Priority tier at a 75% premium and Flex at a 50% discount to Standard. Not every candidate offers both: Qwen3 Coder Next lists Standard only, and none of the cards offers the Reserved tier.

Llama 3, 3.1 and 3.3 70B

For chat, extraction and RAG answers, test Mistral Large 3 (256K context, 32K max output, image input), NVIDIA Nemotron 3 Super 120B (256K context, 32K output, configurable reasoning effort) and DeepSeek V3.2 (164K context, 8K output). Output caps are not a risk: Llama 3 70B caps output at 8K tokens and Llama 3.3 70B at 4K, so every shortlisted model matches or exceeds them.

Llama 4 Maverick and Scout

Mistral Large 3 and Kimi K2.5 accept images and run in-Region in Mumbai; Gemma 4 31B accepts images and video but runs only on bedrock-mantle, under the /openai/v1 path, and not in Mumbai. Qwen3 VL 235B A22B is the larger vision option at $0.53/$2.66 in the US Regions. Watch context length: the Llama 4 cards list 1M tokens for Maverick and 10M for Scout, while these candidates stop at 256K. Kimi K3 offers 1M tokens with image input, at a much higher price.

Small and speech models

AWS prices Ministral 3B, 8B and 14B at flat rates of $0.10, $0.15 and $0.20 per million tokens each way in the US Regions, and $0.12, $0.18 and $0.24 in Mumbai. Gemma 4 E2B and 26B-A4B are priced only in US Regions and Frankfurt. Nemotron Nano 3 30B costs $0.06/$0.24 in the US, against $0.06/$0.23 for the retiring Nemotron Nano 9B v2, though its own earliest end-of-life date is 15 December 2026. The catalogue lists no Voxtral model beyond the two retiring ones, so speech users need a different pipeline or self-hosted weights.

Reasoning and coding models

DeepSeek-R1 and DeepSeek-V3.1 users have DeepSeek V3.2. Qwen3 Coder and Devstral 2 users should test Qwen3 Coder Next (256K context, 16K output). MiniMax M2 and M2.1 users can move to MiniMax M2.5 at the same $0.30/$1.20 US price. Kimi K2 Thinking users can move to Kimi K2.5 ($0.60/$3.00 in the US) or Kimi K3, which AWS prices at $3.00/$15.00 through global cross-Region inference, five times K2.5 on both input and output. GLM 4.7 users have GLM 5 ($1.00/$3.20, in-Region in Mumbai) and GLM 5.3. GLM 5.3 is limited to eligible customers and runs only through US or global cross-Region profiles.

Keep the same weights: Custom Model Import, self-hosting or another provider

If prompts, fine-tunes or evals are tuned to one model, three routes keep the weights.

Bedrock Custom Model Import. It accepts Hugging Face-format weights for Llama 2, 3, 3.1, 3.2 and 3.3, Mistral, Mixtral, Qwen2 and Qwen3 (Qwen3ForCausalLM and Qwen3MoeForCausalLM only) and GPT-OSS, among others. Llama 4, DeepSeek, Gemma, GLM, Kimi and MiniMax architectures are not on the list. Limits: text weights under 200GB, context under 128K, no batch inference, no CloudFormation, and only us-east-1, us-east-2, us-west-2 and eu-central-1. AWS charges $0.05718 per Custom Model Unit (CMU) per minute in US Regions and $0.07144 in Frankfurt, billed in 5-minute windows from the first successful call, plus $1.95 per CMU per month for storage. It quotes 2 CMUs for Llama 3.1 8B with 128K context and 8 CMUs for Llama 3.1 70B with 128K; the exact count is set at import.

SageMaker AI or EC2. Self-hosting gives full lifecycle control and Mumbai residency, at the cost of GPU capacity planning and operations. Our open-weight self-hosting decision guide sets out the break-even arithmetic for a dedicated GPU node.

Another provider. Serverless GPU platforms and other clouds host Llama-class weights on their own retirement schedules; check them, or you trade one deadline for another.

Comparison: four migration paths

Decision table

Worked example: a Mumbai Llama 3 70B workload

Assume a team runs Llama 3 70B Instruct in ap-south-1 at 500 million input and 100 million output tokens a month. The table uses Mumbai Standard-tier prices from the pricing page and its price data. Token counts shift between tokenisers, so measure a sample with each candidate.

Every candidate costs less than the Llama it replaces: Mistral Large 3 comes in about 77% lower ($471 against $2,010), and its Flex tier halves that to $235.50 for work that can wait.

Now Custom Model Import for Llama 3.3 70B in us-east-1, assuming the 8 CMUs AWS quotes for a Llama 3.1 70B model with 128K context:

• Per active hour: 8 × $0.05718 × 60 = $27.45.

• Office hours only (10 hours × 22 days = 220 hours): 220 × $27.45 = $6,038, plus 8 × $1.95 = $15.60 storage.

• Always on (730 hours): 730 × $27.45 = $20,036, plus $15.60.

At this volume, the office-hours import costs about 3 times today's Llama bill ($6,038 ÷ $2,010) and 12.8 times Mistral Large 3 ($6,038 ÷ $471); matching it would take roughly 6.4 billion input and 1.3 billion output tokens a month at Mistral Large 3 rates. It also moves processing out of India. Import weights when they carry value, such as a fine-tune, not to save money.

Quota and capacity changes to plan for

• Per-Region quotas. None of the five Mumbai in-Region candidates offers a geo or global cross-Region profile in commercial Regions, so you cannot burst elsewhere. Request ap-south-1 quota increases before cutover.

• Ramped throughput on bedrock-mantle. AWS says on-demand throughput on that endpoint grows over time, and the Gemma 4 31B card says its default limits do not appear in Service Quotas. Ramp over days, not hours.

• Quota burndown. GLM 5.3 and Kimi K3 count each output token as 10 against quota.

• Feature gaps. None of the eight candidate cards lists Knowledge Bases support, and gpt-oss-120b on bedrock-runtime does not list Agents, Flows or prompt management.

India, EU and US data notes

India. Twenty of the 26 are listed for ap-south-1, including both Llama 3 models, all three Gemma 3 models and both Qwen3 Coder models. In-Region Mumbai options are DeepSeek V3.2, Mistral Large 3, MiniMax M2.5, Nemotron 3 Super 120B and gpt-oss-120b; Kimi K3 adds an India profile that routes only between Mumbai and Hyderabad. GLM 5.3 reaches Mumbai only through a global profile, which can process prompts outside India, and Custom Model Import is not offered there. If your DPDP assessment assumed in-country processing, re-run it whenever the model, endpoint or inference profile changes; our DPDP engineering playbook covers the controls.

EU and UK. London (eu-west-2) appears on 19 of the 26 retirement entries. Mistral Large 3 runs in-Region in London and Stockholm, MiniMax M2.5 and Nemotron 3 Super in several EU Regions, Gemma 4 in Frankfurt, and Custom Model Import in Frankfurt at $0.07144 per CMU-minute.

US. The five Llama 3.1, 3.3 and 4 models and DeepSeek-R1 are listed only in US Regions, so US teams carry most of the Llama exposure.

A 10-week test and cutover plan

Starting in the week of 4 January 2027 puts cutover in the week of 8 March, three weeks before end of life.

1. Week 1: inventory and protect access. Run the four-source audit, record model ID, Region, endpoint, caller and monthly tokens, and add keep-alive calls for any caller idle more than 10 days.

1. Week 2: golden set. Collect 200 to 500 real prompts per use case with expected outputs, edge cases and refusals.

1. Weeks 2 to 3: shortlist. Pick two candidates per use case. Confirm Region, endpoint, max output, Guardrails, managed features and the candidate's own lifecycle state.

1. Weeks 3 to 5: evaluate and fix prompts. Score accuracy, format compliance, tool calls, latency and cost per task, then adjust system prompts, stop sequences, JSON schemas and temperature.

1. Week 6: quotas. File quota requests sized to peak traffic plus headroom.

1. Weeks 7 to 8: shadow traffic. Mirror 5 to 10% of production and compare outputs offline.

1. Week 9: canary. Route a small live share behind a rollback flag.

1. Week 10: cut over. Move all traffic and remove old IDs from code.

A gateway layer, such as those in our AI gateway comparison, turns steps 6 to 8 into a configuration change rather than a release.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. We inventory Bedrock model IDs across accounts and Regions, build golden-set evaluations and move traffic behind a gateway with rollback, as described in our AI gateway and model routing service. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Model lifecycle (Legacy), as published on 7 November 2026 — Amazon Bedrock User Guide, AWS

1. Amazon Bedrock pricing — Amazon Web Services

1. Amazon Bedrock price data file behind the pricing page, published 6 November 2026 — Amazon Web Services

1. FoundationModelLifecycle — Amazon Bedrock API Reference, AWS

1. Models at a glance — Amazon Bedrock User Guide, AWS

1. Llama 3 70B Instruct model card — Amazon Bedrock User Guide, AWS

1. ListFoundationModels — Amazon Bedrock API Reference, AWS

1. Model lifecycle — Amazon Bedrock User Guide, AWS

1. GLM 5.3 model card — Amazon Bedrock User Guide, AWS

1. Kimi K3 model card — Amazon Bedrock User Guide, AWS

1. DeepSeek V3.2 model card — Amazon Bedrock User Guide, AWS

1. Mistral Large 3 model card — Amazon Bedrock User Guide, AWS

1. Qwen3 Coder Next model card — Amazon Bedrock User Guide, AWS

1. MiniMax M2.5 model card — Amazon Bedrock User Guide, AWS

1. NVIDIA Nemotron 3 Super 120B model card — Amazon Bedrock User Guide, AWS

1. Gemma 4 31B model card — Amazon Bedrock User Guide, AWS

1. gpt-oss-120b model card — Amazon Bedrock User Guide, AWS

1. Llama 3.3 70B Instruct model card — Amazon Bedrock User Guide, AWS

1. Kimi K2.5 model card — Amazon Bedrock User Guide, AWS

1. Llama 4 Maverick 17B Instruct model card — Amazon Bedrock User Guide, AWS

1. Llama 4 Scout 17B Instruct model card — Amazon Bedrock User Guide, AWS

1. Nemotron Nano 3 30B model card — Amazon Bedrock User Guide, AWS

1. GLM 5 model card — Amazon Bedrock User Guide, AWS

1. GLM 5.3 by Z.ai is now generally available on Amazon Bedrock, 5 October 2026 — AWS What's New

1. Use Custom model import to import a customized open-source model into Amazon Bedrock — Amazon Bedrock User Guide, AWS

1. Calculate the cost of running a custom model — Amazon Bedrock User Guide, AWS