Meta

Llama 4

Provider

Meta

Status

available

Context

10,000,000 tok

Llama 4 is Meta’s open-weight model family, released on 5 April 2025 — the first open-weight, natively-multimodal models built on a mixture-of-experts (MoE) architecture, and the high-water mark of Meta’s open-source era. The “herd” came in three sizes: Scout (17B active / 109B total, with an industry-leading 10-million-token context window), Maverick (17B active / 400B total, the assistant workhorse), and Behemoth (288B active / ~2T total, a teacher model that was ultimately shelved) (Meta, Hugging Face).

In 2026 Llama 4 is legacy for Meta’s own assistant — Muse Spark replaced it as the Meta AI engine — but it remains freely available for developers to download, run and fine-tune. Its licence is not Apache 2.0: the Llama 4 Community License carries usage obligations and a 700-million-monthly-active-user threshold that requires a separate licence from Meta (Royfactory).

Verdict — updated 26 July 2026. Llama 4 is no longer Meta’s frontier play: Behemoth has been shelved, and Meta’s newest model is the closed-weight Muse Spark, so the Llama line is effectively paused at the top end. That doesn’t make Llama 4 irrelevant, though — Scout and Maverick remain freely downloadable open weights, and Scout’s 10M-token context plus single-GPU quantised builds are exactly why people still reach for it: private, self-hostable, long-context inference you fully control. If you want a local Llama in 2026, see how to run it locally below.

Quick specs

GET LLAMA 4 →

The Llama 4 herd

All use a mixture-of-experts design — one expert is always active for general knowledge, with others selected per token — which is what lets Scout deliver a 10M-token context on accessible hardware (Meta). Scout was pretrained on ~40 trillion tokens and Maverick on ~22 trillion, of multimodal data.

Benchmark performance

At its April 2025 release, Meta reported that Llama 4 Maverick beat GPT-4o and Gemini 2.0 Flash across a range of widely-reported benchmarks and matched DeepSeek v3 on reasoning and coding at less than half the active parameters, while Scout outperformed Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 (Meta, TechTalks).

A credibility caveat: Meta’s then-Chief AI Scientist Yann LeCun later alleged that the GenAI team had “fudged” some Llama 4 benchmark results — an unverified claim, but one that contributed to Zuckerberg’s loss of confidence in the team and the subsequent reorganisation (see the Meta provider page). Either way, the open frontier has since moved well ahead — models like DeepSeek V4 now lead on open-weight capability — so Llama 4’s draw today is its licence-able weights and Scout’s enormous context, not raw benchmark standing. See best AI models.

Licence and access

Llama 4 is open weights, but not Apache 2.0. The Llama 4 Community License allows free download, self-hosting, fine-tuning and commercial use, but with practical obligations and a 700-million-monthly-active-user threshold above which a separate licence from Meta is required (Royfactory) — a meaningful distinction from the fully-permissive licences on gpt-oss, Gemma 4 or DeepSeek.

The weights are on Hugging Face and llama.com, with hosting across the major clouds and inference providers. Scout fits on a single server-grade GPU; Maverick ships in BF16 and FP8.

How to run Llama 4 locally

Llama 4’s biggest draw in 2026 is that you can run it yourself — no API, no data leaving your machine, and Scout’s long context intact. What you need depends on which model and how much quantisation you accept.

Llama 4 Scout — the local-friendly one

Scout (17B active / 109B total, 16 experts) is the one most people actually run at home:

Quickest start: ollama run llama4:scout. Most control: llama.cpp with an Unsloth GGUF quant, which lets you pick the exact bit-width and context length for your hardware. Note the trade-off: the 1.78-bit quant that fits one card is meaningfully weaker than full precision, and pushing toward Scout’s 10M-token context needs far more memory than the weights alone.

Llama 4 Maverick — needs a real rig

Maverick (17B active / 400B total, 128 experts) is not a single-consumer-GPU model. Realistically it wants a multi-GPU workstation — around two 48GB cards — or a cloud instance. It ships in BF16 and FP8; quantised GGUFs exist, but the footprint still rules out most desktops.

Behemoth — you can’t

The ~2T-parameter Behemoth teacher model was shelved and never released, so there are no public weights to run.

Is local Llama 4 worth it?

Running locally buys you privacy, offline use and zero per-token cost, but you pay in setup and capability: heavily-quantised Scout trails the current hosted open leaders like DeepSeek V4. If you want maximum open-weight capability, check the best AI models ranking; if you specifically want a model that runs on one GPU with enormous context, Scout is still one of the strongest picks.

How Llama 4 compares

• vs Muse Spark — Muse Spark replaced Llama 4 as the Meta AI engine and matches Maverick at far less compute, but it is proprietary; Llama 4 is the open option.
• vs other open models — Against DeepSeek, Alibaba’s Qwen, Mistral and Gemma 4, Llama 4’s standout is Scout’s 10M-token context; on raw capability the newer open models now lead, and Llama 4’s licence is more restrictive than Apache/MIT alternatives.

Known limitations

Superseded as Meta’s assistant by Muse Spark. Restrictive licence — the 700M-MAU clause makes it less “open” than Apache/MIT rivals. Benchmark-credibility cloud from the LeCun allegation. And the open frontier has moved ahead since April 2025, so Llama 4 trails the current best open-weight models on capability.

FAQ

What is Llama 4?

Llama 4 is Meta’s open-weight model family, released April 2025 — the first open-weight, natively-multimodal mixture-of-experts models. It comes in Scout (10M context), Maverick (the workhorse) and the in-training Behemoth.

Is Llama 4 free and open source?

It is open weights and free to download, but under the Llama 4 Community License rather than a fully-permissive Apache/MIT licence — there are usage obligations and a 700-million-monthly-active-user threshold that requires a separate licence from Meta.

What is Llama 4 Scout’s context window?

Up to 10 million tokens — the largest of any widely-available model at its release — while fitting on a single server-grade GPU with quantisation.

Is Llama 4 still Meta’s main model?

No. Muse Spark replaced Llama 4 as the engine behind Meta AI in April 2026, and it is closed-weight — Meta’s frontier work has moved off the open Llama line. Llama 4 remains available as open weights for developers.

How do I run Llama 4 locally?

Scout is the practical choice: it fits on a single 24GB consumer GPU using Unsloth’s 1.78-bit dynamic GGUF through llama.cpp (~20 tok/s), or on a 32GB+ Apple Silicon Mac at Q4 via Ollama (~15–32 tok/s). The fastest start is ollama run llama4:scout. Maverick needs a multi-GPU rig (roughly two 48GB cards) or the cloud, and Behemoth was never released. See how to run Llama 4 locally above for the full breakdown.

Was Llama 4 Behemoth ever released?

No. The ~2-trillion-parameter Behemoth was previewed in April 2025 as a “teacher” model but was repeatedly delayed and ultimately shelved — Meta lost confidence that its gains justified shipping, and by 2026 had pivoted to the closed Muse Spark. It was never formally cancelled, but no public weights exist.

Last verified 26 July 2026. Llama 4 figures are Meta-reported from the April 2025 release; the “fudged benchmarks” allegation is attributed to Yann LeCun and is unverified. Local hardware figures are community-reported (Unsloth/llama.cpp/Ollama) and vary by quant and context length. Licence terms and benchmarks change — confirm against Meta’s official pages and the model licence before relying on them.