Model releases · 11 August 2026
Meta is still publishing open weights. They are no longer called Llama, and they are no longer on the account most people are watching.
| Most recent Llama release | Llama 4 (Scout 17B-16E, Maverick 17B-128E), 2025 |
|---|---|
| Most recent Meta open release | Muse Glimmer 30B, August 2026, on a different account |
| Llama 4 architecture | Llama4ForConditionalGeneration, Mixture-of-Experts, multimodal |
| Llama 4 Scout size | 217.3 GB |
| Llama 4 licence | Llama 4 Community License — not Apache or MIT |
| Muse Glimmer licence | Apache 2.0 |
| Llama 3.2 sizes still in wide use | 1B and 3B |
| Accounts | meta-llama for Llama, meta-models for Muse |
There is no Llama 5. Meta's newest open-weight model is Muse Glimmer 30B, published in August 2026 under the Muse name, from the meta-models account rather than meta-llama, and under Apache 2.0 rather than a Llama community licence.
Anyone watching the meta-llama account for the next release would have missed it.
The last models to carry the name are Llama 4 Scout and Llama 4 Maverick, both Mixture-of-Experts and multimodal, at 17B active parameters with 16 and 128 experts respectively. Scout's weights are 217.3 GB. Both ship under the Llama 4 Community License, which carries usage conditions that Apache 2.0 does not.
The account also still hosts Llama Guard 4 and Prompt Guard, the safety classifiers, and the Llama 3 line that a great deal of downstream work is still built on.
Llama 3.2 1B and 3B remain among the most widely deployed models on phones and embedded hardware, and that has little to do with which model is newest. They are small, well understood, supported by every runtime, and quantise predictably. A newer model has to beat all four of those properties, not just a benchmark.
The shift from a Llama community licence to Apache 2.0 on Muse Glimmer is the most consequential part of the rename for anyone building commercially. The Llama licences carried conditions on scale of use and on naming derivative works. Apache 2.0 does not.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | The llama architecture is the runtime's namesake and has the broadest support of any family. |
| MLX | Yes | Conversions available across the Llama 3 line. |
| transformers | Yes | Reference implementations for every generation. |
| Llama 4 in llama.cpp | Partial | llama4 support exists but the multimodal path is less complete than the text path. |
No. Meta's newest open-weight model is Muse Glimmer 30B, released in August 2026 under the Muse name from a different Hugging Face account, under Apache 2.0.
Llama 4, in Scout and Maverick variants, both Mixture-of-Experts and multimodal, released in 2025 under the Llama 4 Community License.
They are small, supported by every runtime, well understood and quantise predictably. Those properties matter more for on-device deployment than being the newest release.
Muse Glimmer 30B ships under Apache 2.0, unlike the Llama models, which carried Meta's own community licences with conditions on scale of use and naming of derivatives.
The meta-models account on Hugging Face, rather than meta-llama.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026