Meta AI and Llama

Models and offerings #

Who they are

This hub is for anyone running or fine-tuning an open-weight model, because for three years that mostly meant Llama, and for anyone trying to follow Meta’s 2026 change of course. Facebook AI Research was founded in 2013, created PyTorch in 2017, and released the first Llama weights to researchers in February 2023. The leak and then the commercial Llama 2 licence in July 2023 made “open weights” a category. Llama 3 (2024) and Llama 4 (April 2025) followed, and the Llama ecosystem (llama.cpp, Ollama, LoRA recipes) grew up around them.

In 2025 Meta reorganised its model work under Meta Superintelligence Labs and in April 2026 shipped Muse Spark, its first frontier model with no downloadable weights, through a new Meta Model API. The open-weight story resumed in October 2026 with Muse Spark Open. Meta is therefore now both an open-weight publisher and a closed API vendor, and the two halves have different licences.

Tooling and APIs

There is nothing to install for the models themselves. Llama checkpoints are gated on Hugging Face behind the licence click-through and load with Transformers, vLLM, llama.cpp and MLX. Meta’s own llama-stack tooling exists but most people use the community stack covered on the open-source hub.

The Llama 4 models, from the April 2025 release: Llama 4 Scout (109 B total, 17 B active across 16 experts, Llama 4 Community License) and Llama 4 Maverick (400 B total, 17 B active across 128 experts, 1 M context). Scout’s model card advertises a 10 M token context, but Meta’s own documentation says it was trained at 256 k and the rest is length generalisation, so treat 10 M as a vendor claim. Both take image input. Llama 3.1 8 B (8 B parameters, 128 k context, Llama 3.1 Community Licence, July 2024) remains the most common fine-tuning base in tutorials, including our LoRA project.

Muse Spark Open (816 B total, 42 B active, 1 M context, Muse Model License, 15 October 2026) is on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B and hosted by Baseten, Fireworks, Groq, Together and others. At 816 B it is a data-centre model, not a laptop one. Meta’s blog calls it the most capable US open-weight model to date. That is the vendor’s claim, and the “Key benchmarks” section of the post gives no scores in the text.

The hosted Muse models are reached with the OpenAI SDK pointed at Meta’s endpoint:

# needs-api
import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["META_API_KEY"], base_url=os.environ["META_API_BASE_URL"])
chat = client.chat.completions.create(
    model="muse-spark",
    messages=[{"role": "user", "content": "In one sentence, what does a learning rate control?"}],
)
print(chat.choices[0].message.content)

Pricing for the Meta Model API is described on the developer site only as “affordable”, and we found no public per-token table, so see dev.meta.ai. Pain points: the gated downloads need a Hugging Face token tied to an approved request, the Llama 4 models are too large for a single consumer GPU without heavy quantisation, and the Muse model names (Spark 1.2, 1.3, Open, Wonder, Wonder Full, Glimmer) have arrived faster than the documentation.

Migration notes

Llama to anything else: Llama’s architecture (RMSNorm, rotary embeddings, SwiGLU, grouped-query attention) became the template that Mistral, Qwen, Gemma and most others follow, so serving stacks treat “Llama-like” as the default path. Conversions to GGUF for llama.cpp and to MLX for Apple silicon are mechanical. The chat template changed between Llama 2, 3 and 4, so prompts built by hand must be regenerated with apply_chat_template.

From a hosted API: a self-hosted Llama or Muse Spark Open behind vLLM gives an OpenAI-compatible endpoint, so application code is unchanged and the migration cost is in operating GPUs and in losing hosted features such as built-in web search. The Mistral hub and the Google hub describe the two other open-weight families to compare against. The models reference lists sizes and licences.

Naming: a derivative of a Llama model must have “Llama” at the start of its name under the Community License, which is why fine-tunes on Hugging Face are called Llama-3.1-8B-something.

Watch for

The licences are the sharp edge. The Llama 4 Community License is not open source: it requires the “Built with Llama” notice, the naming rule above, an acceptable use policy, and a separate licence for any service with more than 700 million monthly active users on the release date. The Llama 4 licence text has also carried a restriction on multimodal use by entities domiciled in the EU, so check the current wording before deploying in Europe. The Muse Model License is described by Meta as “limited commercial terms” and the full text should be read before building a product on the weights. None of these allow you to call the result open source.

Meta’s most capable models (Muse Wonder) are now API-only, and the company has said little about future Llama releases. Plan on the possibility that the next Meta model you want is closed.

Training data for Llama and Muse is not disclosed in detail. Vendor benchmark numbers in Meta’s launch posts are vendor numbers. Consumer Meta AI products train on user interactions in many jurisdictions, which is separate from the developer API. Read the terms for the surface you use.