Llama 4

Overview

Llama 4 is Meta's last open-weight frontier release. Published on April 5, 2025, it introduced Scout and Maverick: two natively multimodal Mixture-of-Experts models that each activate 17 billion parameters per token, and pushed open-weight context windows to 10 million tokens.

Fifteen months later, no new Llama foundation model has shipped. There is no Llama 4.5 and no Llama 5. The line was extended once, at LlamaCon on April 29, 2025, with safety components rather than a base model: Llama Guard 4, Llama Prompt Guard 2, and LlamaFirewall. Llama 4 Behemoth — announced in the same April 2025 post at 288B active parameters and roughly 2 trillion total — was still training and was never publicly released.

Meta's frontier work has moved elsewhere. On April 8, 2026, Meta Superintelligence Labs released Muse Spark, a natively multimodal reasoning model with tool use, visual chain of thought, and multi-agent orchestration. Muse Spark ships with no downloadable weights: it is available through meta.ai, the Meta AI app, and a private API preview for select users.

Meta has published no statement discontinuing Llama. What the record shows is narrower and clearer: the company's frontier releases are now proprietary, and no new Llama foundation model has shipped since April 2025 — only the open-weight safety components released at LlamaCon that same month. Meta also shut down its own hosted Llama API on July 6, 2026.

Model Family

Llama 4 Scout

• Active parameters: 17B
• Total parameters: 109B
• Experts: 16
• Context window: 10,000,000 tokens
• Distinguishing feature: iRoPE — interleaved attention layers carrying no positional embeddings, with rotary position embeddings (RoPE) in most other layers and inference-time temperature scaling of attention; the mechanism behind the 10M window

Meta positioned Scout against Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 at launch.

Llama 4 Maverick

• Active parameters: 17B
• Total parameters: 400B
• Experts: 128
• Context window: 1,000,000 tokens
• Distinguishing feature: the general-purpose workhorse of the pair, stronger on reasoning and coding

Meta's April 2025 claim was that Maverick beat GPT-4o and Gemini 2.0 Flash, with results comparable to DeepSeek v3 on reasoning and coding.

Llama 4 Behemoth — never released

• Active parameters: 288B
• Total parameters: ~2 trillion
• Experts: 16

Previewed as still in training. Meta claimed it outperformed GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks. It served as an internal teacher model for distillation and never shipped as public weights.

Capabilities

• Native multimodality via early fusion: Text and vision tokens are integrated in a unified backbone from the start, rather than through a separately trained vision encoder.
• Mixture-of-Experts efficiency: Alternating dense and MoE layers mean only 17B parameters activate per token, on models totalling 109B and 400B.
• 10M token context (Scout): The longest context window in any widely deployed open-weight model, enabled by iRoPE.
• Massively multilingual pre-training: 200 languages, including more than 100 with over a billion tokens each.
• Open weights: Downloadable and self-hostable under the Llama 4 Community License Agreement.
• Text and code output: From text and image input.

Technical Specifications

• Architecture: Mixture-of-Experts with alternating dense and MoE layers, on a Transformer backbone
• Multimodality: Early fusion of text and vision tokens
• Long context (Scout): iRoPE — interleaved attention layers with no positional embeddings, rotary position embeddings (RoPE) in most other layers, and inference-time temperature scaling of attention
• Training corpus: More than 30 trillion tokens across 200 languages
• Input modalities: Multilingual text and images
• Output modalities: Multilingual text and code
• License: Llama 4 Community License Agreement (custom commercial license)
• Officially supported languages (Maverick Instruct model card): Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, Vietnamese
• Knowledge cutoff: August 2024, per Meta's Maverick Instruct model card

Where Llama 4 Stands Today

Llama 4 is still a capable, freely downloadable multimodal MoE family, and Scout's 10M token window has no open-weight equal. But three things have changed since April 2025:

1. Meta's frontier moved to closed weights. Muse Spark (April 8, 2026) is a Meta Superintelligence Labs model with no public checkpoint.
2. The open-weight field caught up and moved on. Gemma 4 shipped under Apache 2.0 on March 31, 2026 with a 31B dense model Google places at #3 among open models on the Arena text leaderboard, and DeepSeek V4 and Qwen3.7-Max both post newer checkpoints.
3. Llama 4's benchmark claims are dated. Meta's comparisons were drawn against GPT-4o, Gemini 2.0 Flash, and Claude Sonnet 3.7 — models that have themselves been superseded twice over.

Choose Llama 4 today for its license terms, its 10M token context, and its maturity in the local-inference tooling ecosystem — not because it is the strongest open-weight model available.

Use Cases

• Extreme-length retrieval and analysis: Scout's 10M token window handles entire document archives or repository histories without chunking.
• Self-hosted multimodal inference: Text and image understanding on private infrastructure with no per-token cost.
• Fine-tuning base: A well-supported, widely documented starting point for domain adaptation.
• Cost-controlled serving: 17B active parameters keeps inference cost far below what the 109B and 400B totals suggest.
• Multilingual applications: 200 languages in pre-training, 12 officially supported in the instruction-tuned checkpoints.
• Research on MoE and early fusion: Open weights make Llama 4 one of the few frontier-scale MoE models available for architectural study.

Performance / Benchmarks

Meta's April 2025 announcement leads with comparative claims. These comparisons are against 2025-era models and should not be read as current standings:

• Scout outperforms Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1.
• Maverick beats GPT-4o and Gemini 2.0 Flash, with results comparable to DeepSeek v3 on reasoning and coding.
• Behemoth (unreleased) outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks.

One score appears in the announcement itself: Maverick reached an ELO of 1417 on LMArena. Read the qualifier carefully — Meta attributes that number to "an experimental chat version," not to the weights it published. The two are not the same artifact.

Fuller benchmark tables do exist, on Meta's own instruction-tuned model cards rather than in the blog post. See the Maverick Instruct card and the Scout Instruct card for the current figures.

Limitations

• No new foundation model in fifteen months. Nothing since April 5, 2025 but safety components, and Meta's frontier releases are now closed-weight.
• Behemoth was never released. The strongest model in the announcement does not exist as downloadable weights.
• The license is not open source. The Llama 4 Community License Agreement is a custom commercial license, not Apache 2.0 or MIT. Among its restrictions: entities with more than 700 million monthly active users must obtain separate permission from Meta before using the model at all.
• Meta's own hosted API is gone. The Llama API — free, and in public preview its whole life — shut down on July 6, 2026. Requests to it now return a sunset response.
• Knowledge cutoff of August 2024. Nearly two years stale as of mid-2026.
• 10M tokens is a ceiling, not a promise. Scout's window is architecturally supported; quality and cost at extreme lengths are separate questions from whether the model accepts the input.
• Text output only. Llama 4 understands images; it does not generate them.
• Twelve officially supported languages in the instruct checkpoints, despite 200 in pre-training.
• Serving cost of Maverick. 400B total parameters must be resident in memory even though only 17B activate per token.

Pricing & Access

Llama 4 weights are free to download and self-host under the Llama 4 Community License Agreement.

Meta no longer hosts Llama itself. The first-party Llama API, announced at LlamaCon in April 2025 and never taken out of free public preview, shut down on July 6, 2026; requests now return a sunset response, and Meta directs developers to third-party hosts. There is no first-party paid API to migrate to.

Where to get the weights

• llama.com - Official download portal
• Hugging Face - Base and instruction-tuned checkpoints for Scout and Maverick

Hosted inference is available through major cloud providers and third-party model-hosting platforms; pricing varies by provider and Meta publishes no reference rate.

Muse Spark, Meta's current frontier model, is a separate product: available at meta.ai and the Meta AI app, with a private API preview for select users. Meta has not published Muse Spark pricing.

Ecosystem & Tools

• llama.com - Downloads, model cards, and documentation
• Hugging Face - Llama-4-Scout-17B-16E-Instruct and Llama-4-Maverick-17B-128E-Instruct
• Llama Cookbook - Official recipes and fine-tuning examples
• Local runtimes - llama.cpp, Ollama, vLLM, and LM Studio all support Llama 4 checkpoints
• Cloud hosting - Available through major cloud providers and dedicated inference platforms

Community & Resources

• The Llama 4 herd - Meta's April 5, 2025 announcement
• Introducing Muse Spark - Meta Superintelligence Labs, April 8, 2026
• Llama 4 Maverick model card
• Llama 4 Scout model card
• Llama 4 product page
• Llama Cookbook on GitHub