Llama Models: Versions, Licences and What Replaced Them

Meta's Llama line is still downloadable, but it is no longer Meta's newest open release. The latest Llama weights on Hugging Face are Llama 4 Scout (about 109B parameters) and Maverick (about 402B), both from April 2025, under a gated custom licence. In October 2026 Meta released Muse Spark Open, an 816B-parameter model, as its new open-weights flagship. Checked 31 October 2026.

The short version

Sizes come from the safetensors totals in the Hugging Face API, which count every parameter, and from the repo names for active counts. "Active" means parameters used per token in a mixture-of-experts model.

Why the Llama name matters less than it did

For two years "open model" and "Llama" were nearly synonyms. That has changed for three reasons we can document.

First, other labs now publish larger and, by their own cards, stronger open models under simpler terms. Qwen3.8-27B is Apache-2.0, DeepSeek V4 is MIT and gpt-oss is Apache-2.0. Our Qwen vs Llama page covers that head to head.

Second, Meta's own focus moved. The llama.com address now opens Meta's developer site, which leads with the Muse family: Muse Wonder, Muse Spark 1.3, Muse Spark Open and the Muse Code terminal agent. Llama sits in the documentation menu as "Llama Docs" and the GitHub list, not at the top.

Third, the last Llama release is about eighteen months old at the date of writing, which is a long time in this market. Older weights remain useful, since a file on your disk does not expire, but nothing on the Llama side is getting new training.

Muse Spark Open

Muse Spark Open is the first Meta open model we would put in front of a coding team. The Hugging Face card metadata and Meta's developer page give these facts.

The Muse licence in plain words

We could read the licence text, because Hugging Face prints it in the repository metadata. Its main points:

• You may use, copy, modify (including fine-tune), distill, merge, publish, distribute, sublicense and sell, subject to the conditions below.
• You must show "Built with Muse Spark Open" in the user interface, documentation, website or settings menu of each product that uses the model or a derivative. A derivative includes any model trained or improved using the model or its outputs.
• If you provide third parties with model access as a service, and your group gross revenue passes 20 million US dollars in any twelve months, you must notify Meta and sign a separate agreement first. Internal use, research and products that merely use the model as a component are excluded.
• Meta claims no ownership of outputs.
• You must follow an attached usage policy, which among other things restricts military, weapons, critical infrastructure and impersonation uses, and says the model is not intended for people under 18.

The attribution rule is the one to read twice. Kimi K3's licence only asks for attribution above 100 million monthly users, and Apache-2.0 and MIT ask for none in a UI. The Muse licence asks everyone. That is a small cost for a company that is happy to display it and a real one for a white-label product.

How Meta positions it

Meta's page calls Muse Spark Open "the most capable open-weight model from a US lab" and says it is built for coding agents, multimodal perception and multi-agent workflows. Its benchmark table compares it with Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, Kimi K3 and two models we did not cover, on tests including Terminal-Bench 2.1, DeepSWE 1.1, MCP-Atlas and GDPval-AA. The page flattens into text when fetched, so we cannot map each score to its model and column with certainty. We will not quote them until someone has read the table in a browser.

Meta also sells API access through its Meta Model API, which it says works with OpenAI SDK compatible clients, and publishes a cookbook that builds a GitHub agent with OpenCode. Meta's Muse Spark Open page states no per-token price.

Llama 4 Scout and Maverick

Llama 4 introduced mixture-of-experts to the family and added image input. Both are tagged image-text-to-text on Hugging Face.

Their value today is mostly as a known quantity. Hundreds of tools already have Llama 4 recipes, quantised builds and fine-tunes. If you have a pipeline tuned on Llama 4, nothing forces you to move.

The weakness is plain from the dates. Both were published before every model in our open source LLM lineup. Newer cards from Qwen, DeepSeek, Z.ai and Mistral all report agentic coding scores that the Llama 4 cards never covered.

Older Llama: 3.1, 3.2 and 3.3

The Llama 3 line still has the largest installed base. The 70B Instruct model of Llama 3.3 (November 2024, about 70.6B dense parameters) needs about 70 GB for weights at 8-bit by our arithmetic, so one 80 GB GPU with almost no room for cache, two for comfort, or one with 4-bit quantisation. The 1B, 3B and 8B models from Llama 3.1 and 3.2 are the small models that tutorials, phone demos and test rigs default to. Hugging Face's own download counter, which covers the trailing 30 days, shows Llama-3.1-8B-Instruct at roughly 4.7 million, far above any Llama 4 repository, which tells you where the existing tooling sits. That is a popularity fact, not a quality one.

If you need a small English chat or classification model with a very large ecosystem, those remain reasonable. If you need current coding quality, see Gemma for small models or Qwen for 27B-class ones.

Getting the weights

Every Llama and Muse repository is gated with manual approval. In practice that means:

1. Create a Hugging Face account and open the model page.
2. Submit the access form. For Muse Spark Open the form asks for your full legal name, date of birth, country, affiliation, and your organisation's full name with corporate identifiers, and it warns that mistakes can block access and cannot be edited.
3. Wait for approval, then download with a token.

Plan for that delay in CI. A build that pulls from Hugging Face with an unapproved token fails, so cache the weights in your own storage after approval. Some third-party repositories mirror quantised builds, for example MLX and GGUF uploads of Muse Spark Open, but check that a mirror carries the same licence and that you accept its terms.

Hardware

Treat the right-hand column as sizing, not a recommendation. Key-value cache for long prompts comes on top, and mixture-of-experts models need all experts in memory even though few are active. The Ollama vs vLLM page covers the serving tools, and Runpod, Modal and Lambda are the places to rent the GPUs.

Where else you can get Llama

Llama weights are widely mirrored, so most inference hosts carry some version. Check the catalogue of AWS Bedrock, Hugging Face Inference, Cloudflare Workers AI and Replicate for the exact model and quantisation they serve, and for the price. We did not compile current host prices for Llama on this page.

Which coding agents can use them

Any agent that accepts an OpenAI-compatible endpoint can talk to a Llama or Muse model served by vLLM or another engine. Meta's Muse Code terminal agent is built for Muse models and is installed with a curl script from dev.meta.ai. Meta's cookbook also uses OpenCode with Muse Spark. For third-party agents see OpenCode and Cline.

Meta's safety and guard models

Alongside the chat models, Meta's Hugging Face organisation carries classifier models that are easy to overlook. Llama Guard 4 is a 12B model for screening prompts and responses, uploaded in April 2025. Llama Prompt Guard 2 comes in 22M and 86M sizes and targets prompt injection and jailbreak attempts. They are small enough to run beside any chat model, including non-Llama ones, as a pre-filter in an agent pipeline. Their licences are Meta's own and gated like the rest, so confirm the terms before you ship them inside a product. We have not benchmarked them, and Meta's cards are the only source on accuracy, and we could not read them without access.

Fine-tuning and distillation

Many teams pick Llama because they plan to fine-tune it. The Muse licence says so explicitly: modify (including fine-tune), distill and merge are all named as permitted. It also says the attribution and usage policy travel with any derivative, and defines a derivative to include models trained using the outputs of Muse Spark Open. That second clause matters if you plan to generate synthetic training data with it. A model trained on that data inherits the "Built with Muse Spark Open" display rule.

Compare that with Apache-2.0 models, which place no such condition on derivatives, and with DeepSeek's MIT weights. If your product is a fine-tuned model that you sell, licence terms on derivatives are a larger factor than raw benchmark scores. The Llama 4 licence on Meta's site says that if you use the materials or their outputs to create, train or fine-tune an AI model that you distribute, you must include "Llama" at the beginning of its name, and that a licensee with more than 700 million monthly active users in the preceding calendar month must request a licence from Meta. We did not read the Llama 3.3 licence, so confirm its terms before you start.

Moving away from Llama

Switching a working pipeline is mostly a matter of three changes.

• Prompt format. Chat templates differ. Every new model card in this lineup ships its own template or encoding script, and DeepSeek's cards say outright that they provide an encoding folder rather than a Jinja template. Do not reuse a Llama template on another model.
• Sampling. The newer cards recommend temperature 1.0 for reasoning models. Old Llama prompts often ran cooler.
• Reasoning output. Newer models return thinking in a separate field. Parsers written for plain chat completions will drop or mangle it.

Run your own evaluation set on both models and compare cost per correct answer. For a like-for-like on small models, see Gemma models. For the biggest open models, see DeepSeek and Kimi.

Who should pick what

Our take

Llama is now a legacy name with a large ecosystem, and Meta's real open-weights effort has moved to Muse Spark Open. If you already run Llama 4, you lose nothing by waiting. If you are starting now, compare Muse Spark Open with the Apache and MIT alternatives on the same tasks, and weigh the attribution rule against a permissive licence that asks for nothing. Our model comparison tool lists the options side by side.

Product names are used descriptively. We are not affiliated with Meta.