Short answer: Meta has not announced a model called Llama 5, and there is no official release date for one. As of late August 2026, the most recent open-weight generation Meta has published under the Llama name is the Llama 4 family, and the company’s 2026 frontier work has appeared under a different brand entirely. In April 2026 Meta Superintelligence Labs introduced Muse Spark, described as the first in the Muse family of models, which changes the question most people are actually asking. If you are waiting for Llama 5, the more useful question is whether Meta’s next frontier model will carry the Llama name at all.
What Meta Has Actually Shipped
The Llama 4 announcement remains the reference point for the open-weight line. Meta introduced the family as natively multimodal and shipped it as a set rather than a single model:
- Llama 4 Scout, a 17 billion active parameter model with 16 experts, which Meta said fits on a single NVIDIA H100 GPU and offers a 10M token context window.
- Llama 4 Maverick, also 17 billion active parameters but with 128 experts, positioned on performance-to-cost ratio.
- Llama 4 Behemoth, a 288 billion active parameter model with 16 experts, described at announcement as still training and used to distil the smaller two.
Scout and Maverick were released for download on llama.com and Hugging Face. That is the last open-weight generation carrying a Llama version number.
Two things about that release shaped expectations for a successor. The mixture-of-experts design meant the active parameter count stayed small while total capacity grew, which is why a model in that family could target a single accelerator. And the very large context window moved the practical question from “what fits in the prompt” to “what is worth putting in the prompt”, which changed how teams built retrieval systems around it. Any future Meta open-weight model would almost certainly extend both of those choices rather than reverse them.
The Muse Spark Pivot, and Why It Matters
In April 2026 Meta introduced Muse Spark, and the framing in the announcement is the important part. Meta called it the first product of a ground-up overhaul of its AI efforts and the first step on a scaling ladder, developed by Meta Superintelligence Labs rather than the group that shipped previous Llama generations.
The technical description is a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration. Meta also introduced a Contemplating mode that orchestrates multiple agents reasoning in parallel, reporting 58 percent on Humanity’s Last Exam and 38 percent on FrontierScience Research for that mode.
Two details matter more than the benchmark numbers. First, availability: Muse Spark launched at meta.ai and in the Meta AI app, with a private API preview for selected users. That is a product-first, access-controlled launch, which is a different posture from publishing weights for download. Second, the naming. A company that intended the next flagship to be Llama 5 would have called it Llama 5.
Why the Name Probably Changed
Version numbers are product decisions, not technical ones, and three things make a name change more likely than a delayed Llama 5.
The first is organisational. Muse Spark is attributed to Meta Superintelligence Labs, a distinct group from the one that shipped Llama 1 through 4. New organisations rarely inherit an old product line’s numbering, because the numbering carries the previous team’s roadmap with it.
The second is strategic positioning. Llama’s identity was open weights. That identity was an asset while Meta was competing for developer mindshare against closed providers, and it becomes a constraint the moment a company wants to control access to a frontier model. Launching that model under a name with no open-weight expectation attached avoids a public reversal.
The third is the scaling-ladder language in the announcement itself. Meta described Muse Spark as the first step on a ladder with larger models in development. Ladders imply a sequence, and that sequence has already been given a name that is not Llama.
None of this means the Llama line is finished. Meta has continued to reference Llama models across its research output through 2026, and an open-weight line serves purposes a flagship product does not. It does mean that treating “Llama 5” as the inevitable next release is a weaker assumption than it looks.
What to Know Before You Trust a Release Date
Search for this question and you will find pages confidently stating that Llama 5 launched on a specific day in 2026, with parameter counts and context windows attached. Several of those pages disagree with each other by months. Treat all of them as unverified.
Use these checks before believing any model release claim:
- Does an official domain confirm it? For Meta that means ai.meta.com, llama.com or about.fb.com. If the only sources are aggregator blogs, the release did not happen.
- Are weights or an API actually available? A real release has a download page or documented endpoints. Announcements without either are speculation.
- Do the details converge? Genuine launches produce consistent parameter counts and dates across coverage. Fabricated ones produce contradictions.
- Is there a model card? Open-weight releases ship licence terms and a model card. No card, no release.
- Check the publication date against the claim. Pages dated before a supposed release that describe it in past tense are recycling speculation.
Prediction markets are also a reasonable signal for this specific question, because they price the probability of a public release rather than restating rumours, and they settle on verifiable criteria. They are not evidence of anything by themselves, but a market pricing a release as unlikely is a useful counterweight to a confident blog post claiming it already happened.
It is worth understanding why this particular pattern is so persistent. Model release queries spike long before any release exists, so the incentive is to publish first and correct later. Corrections rarely arrive, because the page keeps earning traffic either way. The result is a search results page where the most confident sources are the least reliable ones, which is the inverse of what readers assume.
What a Future Llama Release Would Plausibly Include
This is informed extrapolation from Meta’s published direction, not inside information. Based on what the Llama 4 announcement prioritised and what Muse Spark demonstrates, the next open-weight Meta model would most likely emphasise:
| Direction | Evidence from published work | Why it matters to you |
|---|---|---|
| Mixture-of-experts architecture | Llama 4 shipped 16 and 128 expert variants | Lower serving cost at a given capability |
| Very long context | Scout shipped with a 10M token window | Whole-repository and whole-corpus workflows |
| Native multimodality | Both Llama 4 and Muse Spark are multimodal from the ground up | Fewer separate vision pipelines |
| Agentic tool use | Central to the Muse Spark description | Models that act rather than only answer |
| Parallel reasoning modes | Contemplating mode orchestrates multiple agents | Higher accuracy on hard tasks at higher cost |
What is genuinely uncertain is licensing. Meta’s open-weight strategy was a deliberate competitive position, and the shift to a product-first launch for its frontier model is the strongest available signal that the strategy is being re-examined. Anyone telling you confidently that the next Meta flagship will have open weights is guessing.
How to Plan Without a Release Date
Waiting for a specific model is almost always the wrong build strategy, and it is especially wrong here.
- Build against an interface, not a model. Keep prompts, tool definitions and evaluation separate from the model call so swapping providers is a configuration change.
- Write your evaluation set first. Twenty to fifty real examples with expected outcomes tell you within an hour whether a new model is better for your use case. Benchmarks will not.
- Assume open weights stay available from somewhere. Even if Meta changes direction, the open-weight ecosystem now has several serious contributors. Portability protects you.
- Track cost per successful task, not per token. A cheaper model that fails more often is more expensive.
- Do not rebuild for every release. Set a rule such as reviewing your model choice quarterly or when your evaluation shows a 10 percent gap, and ignore the rest of the news cycle.
A concrete version of this: a small team running a document-processing pipeline kept a fifty-example evaluation set with graded expected outputs. When a new model appeared, one engineer spent ninety minutes running the set against it, compared accuracy and cost per document against the incumbent, and either switched or did not. Over four model generations that team changed providers twice and ignored six releases entirely, with no rebuild in either direction.
The teams that handled the last three model generations well were not the ones who predicted releases correctly. They were the ones whose systems could absorb a new model in an afternoon.
Building the Skill That Outlasts Any Model Name
There is a reason release-date questions attract so much traffic: the pace of releases makes people feel they are falling behind by default. The practical fix is not tracking announcements more closely. It is getting fluent enough with the current generation that the next one is a substitution rather than a relearning exercise.
That fluency is mostly about the parts that do not change between versions: how to structure a prompt, how to build an evaluation, how to decide which tasks to delegate, and how to catch the failure modes that every model shares. Learning those in a structured sequence is far faster than assembling them from release notes, and it leaves you with transferable judgement instead of product-specific trivia. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
FAQ
Is there an official Llama 5 release date?
What is the newest Meta model?
Will the next Meta model be open-weight?
Why do so many sites list a Llama 5 release date?
Should I wait for the next Meta model before building?
How can I access Meta’s newest model now?
What to Do Instead of Waiting
Set a monthly reminder to check ai.meta.com directly rather than reading aggregator coverage, and spend the time you would have spent following rumours on building an evaluation set for whatever you are actually trying to do. When the next Meta model does arrive, under whatever name, that evaluation set will tell you in one afternoon whether it is worth switching, which is a question no release-date article can answer for you.