Llama 5: The Open Model Everyone Keeps Expecting — What We Know So Far

The name has never appeared on an official Meta page, yet it keeps surfacing in the comments under every Meta open-weights post: Llama 5. Recently it showed up in a two-word verdict from a machine-learning researcher on X whose early takes get passed around — @nrehiew_ wrote "Llama 5 basically. Super excited." It is easy to see why the name is in the air: Meta has just released the open weights for Muse Glimmer and its CEO has published a 6,500-word manifesto promising "even larger models." Meta has not announced anything called Llama 5. It has not confirmed such a model exists. And yet the name keeps coming up — in analyst notes, in the comments, and now in an insider's excited two words. This is what the signal actually says, what it does not say, and what a reader building on Meta's models should make of it.

A sourcing note up front, because this story is mostly signal-to-noise. Everything below about a model called "Llama 5" is unverified: no weights, no announcement, no benchmark anyone outside Meta has run. The one thing that is verifiable is that Meta has not shipped it — and that several web pages claiming it shipped in April are wrong. Where a claim comes from a report rather than a spec, it is labeled as such. The models that are real — Llama 4 Maverick, Muse Glimmer, Muse Spark 1.2 — get their own pages; the ghost named Llama 5 gets none.

What the signal actually is

The post is from @nrehiew_, a machine-learning researcher on X who posts technical threads on model papers and internals. "Llama 5 basically. Super excited." That is the entire signal — no screenshot, no spec, no date. The word "basically" is doing a lot of work: it reads like a reaction to something that just arrived, not an announcement of something that will. Timed against the same week's news, the most natural reading is that the researcher looked at Meta's open-weights moves and saw the open flagship the community has been expecting for two years arriving under a different name.

Three readings of "basically Llama 5" are plausible, and the tweet does not discriminate between them. First, the open-weights release that just happened — Muse Glimmer — is, to an insider, the successor everyone kept calling "Llama 5," renamed. Second, a larger open release teased as coming "soon" is being called Llama 5 in advance. Third, the researcher knows something specific about an actual Llama 5 that has not been announced. All three are speculation. One person's two words is enthusiasm, not a roadmap.

The state of the Llama line: verified

Here is what is actually true. Meta's last major Llama release was the Llama 4 family in April 2025 — the open mixture-of-experts line with Scout and Maverick, plus Behemoth, which was previewed and never shipped. A Llama 4.1 followed in early 2026. Then the organization restructured into Meta Superintelligence Labs, and the open line went quiet. An independent analysis published in June 2026 was blunt about the result: as of that month, Meta's flagship open-weight model was still Llama 4 Maverick, there is "no Llama 4.5 and no Llama 5," and teams were advised not to plan around Behemoth or Llama 5 because neither has shipped.

The official pages agree. Meta's own model site now leads with the Muse family — Muse Glimmer and Muse Spark 1.2 — and lists the Llama 4 lineup behind them. It does not list a Llama 5. One of the third-party "Llama 5 spec sheets" that surfaced in search results now returns a 404, which is its own kind of evidence.

The codename that wasn't: Avocado

The "Llama 5" name has been tangled up with a codename since late 2025. Reporting at the time said Meta's next flagship — internally called "Avocado" and widely assumed to be Llama 5 — might not be open-weights at all, and that Meta was weighing closing the line after Llama 4's mixed reception. The target slipped from the end of 2025 into 2026.

What happened next is where the reporting splits, and the split matters. Some outlets report that the Avocado model became Muse Spark, the closed flagship Meta shipped in April 2026 — meaning the "Llama 5" codename effectively died into a proprietary product. Others still treat "Llama 5 / Avocado" as a future open release. One of the more careful open-model trackers, updated in July 2026, puts "Llama 5, codename Avocado" as most likely a 2027 release, citing Goldman Sachs and Finterra forecasts and a consensus below 20% odds on a 2026 ship, with EU AI Act compliance and data availability as the drags. That is the honest state of the rumor: a codename that may already have been consumed by a closed model, and a name that analysts still expect to see on an open one.

The Muse pivot is the real news

The reason "Llama 5" keeps coming up in the same week as Muse Glimmer is that the open line has been renamed, not killed. On August 10, Meta released the weights for Muse Glimmer, a 30-billion-parameter dense multimodal model under the Apache 2.0 license, distilled from the larger Muse Spark 1.2 foundation model and built to run always-on agent workloads on a single consumer GPU — quantized to under 20 GB, inside a 24–32 GB envelope. Zuckerberg's essay the same day committed Meta to open-weights releases, defended distillation, promised that "even larger models" are coming, and said the weights for Muse Spark 1.2 — Meta's most capable closed model — will be open-sourced "in the coming weeks."

So the open flagship the community has been expecting since Llama 4 is arriving in pieces, under the Muse name: a small open model today, the big closed model's weights soon, and larger ones after. None of it is called Llama 5. The question the leak raises is whether any of it ever will be.

The April "release" pages are wrong — here's how you can tell

Searching "Llama 5" surfaces pages that claim Meta released Llama 5 on April 8, 2026, alongside Muse Spark, complete with specs — a 600-billion-parameter mixture-of-experts, a multi-million-token context window, MMLU-Pro scores in the high 80s, per-million-token prices. None of that is corroborated. Dated, independent reporting from June and July 2026 — after that supposed release — explicitly says no Llama 5 exists, that Meta's flagship open-weight model remains Llama 4 Maverick, and that anyone giving Llama 5 a release date is guessing. The spec sheets appear to have been built from each other and from a launch date that contradicts the record. Treat every number on them as unverified. And on Hugging Face, where an open Llama 5 would have to live, no such model is listed.

What it changes for builders

For anyone building on Meta's models, the practical lesson is the same as for every leak: do not plan a production path around an unannounced model, whatever the hype. The testable open tier today is Muse Glimmer (out now) and the promised Muse Spark 1.2 open weights (weeks away, per Meta). Those are real, downloadable, and verifiable. "Llama 5" is not — it is a name in search of a model.

When the big open flagship does land, under whatever name, the safe way to adopt it is the same pattern that works for every new model: point a slice of traffic at it behind automatic failover, so a surprising failure mode costs you a test path rather than a customer-facing one. That is the argument for a router in general — one API that can reach a new model the day a provider hosts it, with provider list prices passed through at 0% markup so a vendor price change is live the same day, and a routing DSL to compose the new model with the ones you already call. The Llama 5 rumors are not on OrcaRouter today, because the model does not exist; no one should pretend otherwise. But the models that actually ship this month already are reachable that way — DeepSeek V4 Pro and DeepSeek V4 Flash are both live on OrcaRouter on one key, with the same pass-through pricing and per-request failover.

What to watch

• The Muse Spark 1.2 weight drop — the single most concrete near-term date in this story. If it ships and is genuinely open, that is the real "Llama 5" moment regardless of what it is called.

• Whether "Llama" survives as a brand — Meta's official site and the manifesto point to Muse. A page that still says Llama is a sign the name lives on.

• A Hugging Face repo from the Meta org for anything larger than 30 billion parameters — the tell that the "even larger models" promise is materializing.

• Independent benchmarks — the vendor numbers on Muse Glimmer are unaudited; third-party evals are what will decide whether the open flagship is actually frontier.

• Corroboration of the signal — the insider post is one person's two words. Other researchers chiming in is how a leak becomes a story.

FAQ

Is Llama 5 already released?

No. Meta has not announced or shipped any model called Llama 5. Pages claiming an April 8, 2026 release are contradicted by dated reporting from June and July 2026, which states plainly that no Llama 5 exists and that Meta's flagship open-weight model remains Llama 4 Maverick.

Will Llama 5 have open weights?

Unknown, and the honest answer is that the question may be moot. Meta's open line is now the Muse family: Muse Glimmer is Apache 2.0 today, and Meta has committed to open-sourcing the weights of Muse Spark 1.2 "in the coming weeks." If the open flagship ships under the Muse brand, there may never be a model literally named Llama 5 — only the open-weights flagship everyone expected, under a new name.

When will it come?

No official date exists. Analysts surveyed by one open-model tracker put a model called "Llama 5" most likely in 2027, with a consensus below 20% odds on a 2026 release. The concrete near-term date to watch is the Muse Spark 1.2 open-weights drop, which Meta says is weeks away.

The pattern in this corner of the market keeps repeating: a name everyone expects, a codename that wanders, a release that arrives under a different label. "Llama 5" may never ship as a product name. What will ship — what already is shipping — is the open flagship Meta's users have been asking for since Llama 4, in fragments: a small open model today, a big model's weights soon, and larger ones after. The insider's two words are excitement about that, not a spec sheet. Read them that way, watch the weight drops, and do not plan around a name.