News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgChameleon

Chameleon

Early-fusion models that read and write images and text in one stream

Chameleon is a family of mixed-modal foundation models announced in May 2024 by the Chameleon Team, with Armen Aghajanyan as the submitting author. They handle images and text together in a single token stream, and the authors report strong results on captioning, text-only tasks and long-form mixed output.

Share on WhatsApp

  • Announced 16 May 2024; revised version dated 21 March 2025
  • Fuses images and text from the start as one token stream
  • Authors report captioning results ahead of prior systems
  • Text-only scores reported above Llama-2, near Mixtral 8x7B and Gemini-Pro
  • Human raters placed long-form mixed output at or above Gemini Pro and GPT-4V

What it is

Chameleon is a group of foundation models that treat pictures and words as one combined stream. The Chameleon Team set out the work in a paper first posted on 16 May 2024, revised in March 2025. Rather than bolting a vision system onto a language model, these systems fuse both kinds of input from the very start, so a single model can read and produce either. The team reports that the family handles captioning, text-only work, picture making and documents that mix both media. Sizes and parameter counts are not given in the available description.

How it came about

The project was presented through an arXiv paper, with Armen Aghajanyan as the submitting author and the work credited to a Chameleon Team. Version one appeared on 16 May 2024 and a second version followed on 21 March 2025. The write-up sets out a training method designed to stay stable from the beginning, a recipe for alignment, and a parameterisation shaped for early fusion. No earlier or later model in the line is named.

Inside the model

The core idea is early fusion: rather than processing images and text separately and joining them later, everything becomes tokens in one shared sequence. The authors say they had to work out a training approach that stays stable from the start, since mixing modalities this way can be fragile. They also describe an alignment recipe, meaning steps taken after pre-training to steer behaviour, and an architectural parameterisation fitted to this setting. Details such as layer counts, attention layout, tokenizer design and training data are not given in the available description.

What it can do

The team evaluated the models across a broad set of tasks. They report leading results on image captioning, and say text-only performance beats Llama-2 while staying comparable with Mixtral 8x7B and Gemini-Pro. Picture generation is described as non-trivial, all within one model. On a new evaluation of long-form output mixing images and text, human raters placed it at or above far larger systems, including Gemini Pro and GPT-4V, according to the authors.

How to use it

The work is published as a research paper on arXiv. The available description does not state whether weights are released, nor does it name a licence, an app or an API.

Why it matters

The authors present Chameleon as a move towards modelling whole documents that mix media within one system, instead of keeping separate models for each kind of content. Beyond that claim, the available description offers no account of later work building on it.

Questions people ask about Chameleon

What makes Chameleon different from other multimodal models?

It uses early fusion, turning images and text into one shared token sequence from the outset. Many other systems process each type of content separately and combine the results later. The authors argue this lets one model handle arbitrary mixes of pictures and words.

Who built it and when was it released?

A group called the Chameleon Team, with Armen Aghajanyan submitting the paper. The first version appeared on 16 May 2024, and a revision followed on 21 March 2025. No company or lab is named in the available description.

How large are the models?

The available description does not give parameter counts, layer numbers or training data size. It refers to a family of models without listing individual sizes.

How does it compare with other systems?

The authors report leading image captioning results, text-only performance above Llama-2 and on a par with Mixtral 8x7B and Gemini-Pro, and human ratings at or above Gemini Pro and GPT-4V on long-form mixed output. These are the team's own reported figures.

Can I download or run it?

The available description covers a research paper only. It does not state whether weights are available, nor does it mention a licence, hosted app or API.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Chameleon yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2405.09818. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.