Llama 5: Release Date, What Meta Has Confirmed and What to Run Instead

Llama 5 has no release date because Meta has not announced it. On 14 November 2026 the newest Llama models on Meta's Hugging Face organisation were still Llama 4 Scout and Maverick from April 2025. Meta's new models use the Muse name instead, and two have open weights: Muse Glimmer, a 30B model under Apache 2.0, and Muse Spark Open, an 816B model under Meta's own licence.

People search for Llama 5 as if it were on a calendar somewhere. It is not. What exists is a set of Meta statements about open models, two open-weight releases under a different brand, and a layer of articles that reported a Llama 5 launch which did not happen.

This page sorts those into confirmed, reported and rumoured, each with a date and a source. We read Meta’s developer site, its announcements, Mark Zuckerberg’s August essay and both Meta organisations on Hugging Face on 14 November 2026. Then it covers the practical questions: which Llama models you can still download, what hardware Meta’s newer open models need, and how to check the next “Llama 5 is out” headline in under a minute.

Llama 5 release date: there is none

Three checks, all repeatable, support that answer.

1. Hugging Face. The meta-llama organisation lists no repository with Llama 5 in its name. Sorted by creation date, the newest entries are Llama Prompt Guard 2 (28 April 2025), Llama Guard 4 (23 April 2025) and the Llama 4 Scout and Maverick checkpoints (1 to 4 April 2025).
2. Meta’s code repository. The models folder of meta-llama/llama-models stops at llama4, and the repository’s last commit was on 10 October 2025.
3. Meta’s developer site. The address llama.com now returns a permanent redirect that ends at dev.meta.ai. That site’s Llama documentation covers Llama 4, Llama 3 and Llama 2. Apart from SAM 3.1, its model list is all Muse, and the hosted product is called the Meta Model API, described as access to Muse models.

So the honest release date for Llama 5 is “not announced”. Anything more specific needs a Meta source behind it, and we found none.

Llama 5 confirmed, reported and rumoured

The state of each claim on 14 November 2026:

The April 2026 rumour has a simple origin. Meta did make a model announcement on 8 April 2026, and it was Muse Spark, not Llama 5. Several sites published “Meta releases Llama 5” stories around that date with parameter counts and context windows that did not agree with each other. One of them, RAGyfied, now carries a correction table stating that its earlier claim of an 8 April Llama 5 release with downloadable weights was wrong and that Meta has not announced Llama 5.

Meta Llama 5 or Muse: what Meta has shipped since Llama 4

The timeline explains why the question keeps coming up.

Between those entries, Meta’s developer site also lists hosted-only models: Muse Spark 1.1, 1.2 and 1.3, Muse Wonder and Muse Ember. None has downloadable weights.

A year passed between Llama 4 and the next Meta model, and that model was closed. The Muse Spark announcement says it is in “private preview via API to select partners” and that Meta hopes “to open-source future versions of the model”. That gap is where Llama 5 speculation grew.

The two open releases that followed sit at opposite ends of the size range, and neither is a like-for-like sequel to Llama 4:

• Muse Glimmer is the small one. Meta’s announcement calls it an open agentic model “optimized for always-on local agent workflows” and small enough for a Mac or a PC with a single consumer GPU. Its model card says it was distilled from Muse Spark, was trained on data from more than 100 languages, supports four reasoning-strength settings from low to xhigh, and has a knowledge cutoff of 4 January 2026.
• Muse Spark Open is the large one. Meta’s developer page describes it as derived from the Muse Spark model released in August, with additional safety post-training, and calls it “the most capable open-weight model from a US lab”. That is Meta’s own description, not an independent ranking. Meta lists the weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, and hosted access through providers such as Fireworks AI, Groq, OpenRouter and Together AI.

A hosted endpoint for an open-weight model is still a hosted endpoint: prompts and files go to that provider. Only the downloaded weights on your own hardware keep data inside your network.

Is Llama 5 open source? What Meta has said about open weights

Meta has said nothing about Llama 5, open or closed. It has said three relevant things about open models in general.

• 8 April 2026. The Muse Spark post expresses a hope to open-source future versions. It makes no commitment and names no date.
• 10 August 2026. In an essay titled The future is for everyone, signed “Mark”, Zuckerberg writes: “Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon.” The essay does not mention Llama or Muse by name.
• 10 August and 15 October 2026. Meta released Muse Glimmer and then Muse Spark Open with downloadable weights.

Two cautions for anyone reading “open” as “open source”. Muse Glimmer’s Apache 2.0 licence is a standard permissive licence. Muse Spark Open’s is not: its model card describes a Muse Model License that requires products to display “Built with Muse Spark Open”, binds users to a usage policy, and requires model-as-a-service providers with more than $20 million in gross revenue over twelve months to notify Meta and sign a separate agreement. Access on Hugging Face is approved manually, after a form that asks for a legal name, date of birth and organisation details. We could not read the full licence text without that approval, so check it yourself before relying on this summary.

The older Llama licences are custom too. The Llama 4 Community License adds commercial terms above 700 million monthly active users, and its use policy withholds the licence for multimodal Llama 4 models from individuals and companies based in the European Union. Our review checklist for open-weight licences lists what to ask counsel. This is general information, not legal advice.

Llama models you can download today

If you specifically want a Llama-branded model, these are the current options. Sizes are 4-bit files as published on 14 November 2026; the “fits” column is our estimate with room for context.

At full 16-bit precision, Ollama lists Scout at 217 GB and Maverick at 803 GB. Both Llama 4 models are mixture-of-experts designs that activate 17B parameters per token, so they generate faster than their file size suggests but still have to be loaded in full.

To run the smallest of them:

ollama run llama3.1:8b

Meta’s Llama organisation also still hosts the small Llama 3.2 models from September 2024, at 1B and 3B for text and 11B and 90B with vision, and two safety classifiers from April 2025, Llama Guard 4 12B and Prompt Guard 2. Those classifiers are the last repositories Meta added under the Llama name.

Higher-precision files cost proportionally more memory. Llama 3.3 70B is 74.98 GB at Q8_0 against 42.52 GB at Q4_K_M in Unsloth’s repository, and Ollama’s 8-bit Scout build is 117 GB against 67 GB. For most single-GPU set-ups the 4-bit file is the practical one.

For Llama across a whole organisation, including the EU clause, serving stacks and sizing for many users, see our enterprise guide to deploying Llama on your own servers. This page does not repeat it.

Llama 5 local requirements: what to plan for

You cannot size hardware for a model with no parameter count. You can size it for the two open models Meta has actually released in 2026, which bracket the range a future release might fall in.

The small end: Muse Glimmer 30B

The Muse Glimmer model card describes a dense model of about 29.6B parameters, including a 1.8B vision encoder, with text and image input and a context length of 131,072 tokens or more. Meta publishes its own quantised builds:

The degradation figures are Meta’s average over 15 benchmarks. The release also includes a 1.63 GB “DFlash” drafter for speculative decoding. With the 17 GB build and the drafter, Meta measured 233.4 tokens per second on an RTX 5090 in llama.cpp against 74.9 without speculation, and 50.2 against 26.6 on an M5 Max using ExecuTorch. Those are Meta’s numbers at batch size 1 with greedy decoding, not ours.

On Ollama the model is muse-glimmer, with an 18 GB default tag. If the quantisation labels are unfamiliar, our guide to reading GGUF quantisation levels explains Q4_K_M and its neighbours.

The large end: Muse Spark Open

The Muse Spark Open model card describes a mixture-of-experts model with 816B total parameters, 42B active per token and a 1M-token context, taking text, image, video and audio input. The BF16 repository is about 1,644 GB and the NVFP4 repository about 533 GB. The card states at least 700 GB of VRAM for NVFP4 and 2 TB for BF16, and says the model is optimised to run on two Mac Studios with M5 Ultra using MLX; Meta’s announcement notes that each can address up to 512 GB. Our Mac Studio M5 Ultra guide covers that machine.

What that means for a purchase

• A 24 or 32 GB GPU runs the small open Meta model today, and any 27B to 32B-class successor at 4-bit.
• A 128 GB unified-memory machine or a 96 GB GPU covers Llama 4 Scout and models of roughly 100B parameters at 4-bit.
• Anything in the Muse Spark Open class is a cluster or a pair of 512 GB Macs, not a workstation.

Check a specific model and context length in the VRAM calculator before you buy. Do not buy against a rumoured Llama 5 size.

How to check a Llama 5 claim yourself

Release claims are cheap to test. A real Meta open-weight release leaves three traces: a post on a Meta domain, a repository in a Meta organisation on Hugging Face, and a licence for that exact model.

This command lists the five newest repositories in the Llama organisation with their creation dates:

curl -s "https://huggingface.co/api/models?author=meta-llama&sort=createdAt&direction=-1&limit=5" \
  | python3 -c "import json,sys; [print(m['id'], m['createdAt'][:10]) for m in json.load(sys.stdin)]"

On 14 November 2026 the first line it printed was meta-llama/Llama-Prompt-Guard-2-86M 2025-04-28. Change meta-llama to meta-models to see the Muse releases instead; there the newest entries were the two Muse Spark Open repositories.

Then apply three rules:

1. No Meta URL, no release. A launch article that links only to other articles is repeating a claim.
2. Check the publisher of any file. Community hubs accept any repository name. A file called “Llama 5” from an unknown account is not a Meta model.
3. Read the licence for that model. Apache 2.0, the Llama Community Licences and the Muse Model License grant different rights, and the brand name tells you nothing about which applies.

Should you wait for Llama 5?

No, for three reasons that do not depend on guessing Meta’s plans.

There is nothing to wait for yet. A wait makes sense when a vendor has given a date or at least a size. Meta has given neither, and its 2026 releases suggest that a future open model from Meta is more likely to carry the Muse name than the Llama one.

The newest Llama models are ageing. Llama 4 dates from April 2025 and Llama 3.3 from December 2024. Muse Glimmer’s model card compares it with Gemma4-31B and Qwen3.6-27B rather than with any Llama model. In that table, which is Meta’s own evaluation, Muse Glimmer scores 76.0 on SWE-Bench Verified against 77.2 for Qwen3.6-27B and 66.6 for Gemma4-31B. Treat vendor tables as a starting point and test on your own tasks, but the comparison set shows which models Meta itself regards as the current competition.

Switching later is cheap if you prepare for it. Model files are replaceable. Prompts, evaluation sets, access rules and audit records are the parts that take time, and none of them is tied to a model name. Build those around the model you can run today.

The one case for holding off is a hardware purchase sized for a specific rumoured model. Buy for a class of model, such as 30B at 4-bit on a 24 or 32 GB card, and not for a name.

What to run now instead of waiting for Llama 5

Waiting has a cost and, so far, no end date. The open-weight field moved a long way while Llama stood still.

Beyond Meta, the choice is wide. Our model picks for each VRAM tier rank what fits 8 to 32 GB cards. The same “announced or not” exercise for Alibaba’s next generation is in our Qwen 4 status page. For other large open families, see the Kimi K3.1 Pro hardware guide, the DeepSeek V4.1 Flash sizing guide and the Gemma 4 deployment guide.

One practical point: if your applications call models through an OpenAI-compatible endpoint, the model behind it can change without the application changing. Choosing a good model now does not lock you out of a better one later, whatever Meta calls it.

How VDF AI fits

VDF AI does not host, resell or bundle Llama, Muse Glimmer, Muse Spark Open or any other model named on this page, and it has no relationship with Meta. Organisations that approve an open-weight model serve it on their own infrastructure; VDF AI Chat names Llama among the open-weight families teams commonly use, but whether a given Llama or Muse model fits your licence and hardware is your call. VDF AI is a platform for governed, private AI that organisations run on premises, air-gapped or in its cloud, with governed agents, role-based access control and audit logs around whichever models they approve.

The question here, wait for a model that may never ship or standardise on one that exists, is a routing question once more than one team is involved. VDF AI Router registers Ollama and custom on-premises deployments beside the cloud models your policy allows, prefers the local ones when they are available, and has an air-gap mode that routes to local models only. A new open-weight release then becomes one more endpoint to register and evaluate.