Llama 4

Mon, 01 Jan 0001 00:00:00 +0000

Llama 4 is Meta’s fourth and, as of September 2026, newest open-weight generation, released 5 April 2025. It was the first Llama built as a mixture of experts and the first to take images through the same pathway as text. It is also the first generation with no research paper, and the only one that shipped incomplete.

Meta announced three models. Two exist:

Model	Active	Total	Experts	Status
Scout	17B	109B	16	Shipped
Maverick	17B	400B	128 routed + 1 shared	Shipped
Behemoth	288B	~2T	16	Never released

Total is what must be held in memory; active is the share each token is actually multiplied by. Both shipped models activate the same 17 billion parameters per token and differ in how much capacity sits behind that — which is the mixture-of-experts trade stated as plainly as a lineup ever states it.

Llama 3

Mon, 01 Jan 0001 00:00:00 +0000

Llama 3 is Meta’s third open-weight generation, released from April 2024 in 8B, 70B and — with Llama 3.1 in July — 405B sizes. It matters less for what it scored than for what came with it: The Llama 3 Herd of Models, a technical report of a detail and honesty that no lab has matched since, including Meta itself.

It is the last Llama generation with a paper. Llama 4 shipped without one.

Meta on Bitsy Wiki

Llama 4

Llama 3