AI Radar

Meta models

Meta Platforms, Inc., named Facebook, Inc. until October 2021, runs apps including Facebook, Messenger, Instagram and WhatsApp. Its FAIR research team dates from late 2013. Meta says it has released its Llama foundation models openly and expects to train both open and closed models; Meta Superintelligence Labs builds the Muse series of large language models. Known as Facebook, Inc. until 28 October 2021.

Chronology

  1. 2026

    5 models
    1. Released
      Muse · Language, Multimodal, Reasoning Muse Spark 1.3

      Muse Spark 1.3 is a multimodal reasoning model in Meta's Muse series, offered in the Meta Model API and the Muse Code terminal agent. Meta says it improves agentic and coding work and is easier to…

      • Adds a max reasoning effort level beyond xhigh, available on the Standard tier only.
      • Meta says that in comparisons by its engineers it took fewer turns, tool calls and tokens and was less verbose than Muse Spark 1.2 in coding work.
      Primary source Available · no news linked
    2. Released
      Muse · Language, Multimodal, Reasoning Muse Glimmer

      Muse Glimmer is a 30-billion-parameter open-weights model in Meta's Muse series, released on Hugging Face under the Apache 2.0 licence for self-hosting. It accepts text and images and reasons before…

      • First Muse model with published weights, under the Apache 2.0 licence; the Muse Spark models are offered only as hosted services.
      • A dense 30B model that Meta designed to run on a Mac or PC with a single consumer GPU.
      Primary source Available · 1 news item
    3. Released
      Muse · Language, Multimodal, Reasoning, Code Muse Spark 1.2

      Muse Spark 1.2 is a multimodal reasoning model in Meta's Muse series that Meta describes as a coding-focused update to Muse Spark 1.1. It was released with Muse Code, a terminal coding agent in beta…

      • Meta says it scaled up training compute on coding tasks, with more varied training environments and long-horizon tasks such as whole-repository generation.
      • Co-trained with the Muse Code agent harness so that it works with that agent's tools, goals, compaction and subagents.
      Primary source Available · no news linked
    4. Released
      Muse · Language, Multimodal, Reasoning Muse Spark 1.1

      Muse Spark 1.1 is a multimodal reasoning model in Meta's Muse series that Meta says is built for agentic tasks. It succeeded the first Muse Spark in Meta AI and launched with the public preview of…

      • First Muse model offered to developers through a public API: the public preview of the Meta Model API launched with it.
      • Adds trained computer use across applications and orchestration of parallel subagents, which Meta says makes complex projects faster than with Muse Spark.
      Primary source Available · 1 news item
    5. Released
      Muse · Language, Multimodal, Reasoning · Milestone Muse Spark

      Muse Spark is the first model in Meta's Muse series, built by Meta Superintelligence Labs. Meta describes it as a natively multimodal reasoning model with tool use; it launched in the Meta AI app…

      • Meta says a rebuilt pre-training recipe reaches the same capabilities as Llama 4 Maverick with over an order of magnitude less training compute.
      • Trained with reinforcement learning to reason before answering, with a penalty on thinking time; a Contemplating mode has several agents reason in parallel.
      Primary source Status unknown · no news linked

  2. 2025

    3 models
    1. Announced, release date unknown
      Llama 4 · Language · Milestone Llama 4 Behemoth

      Llama 4 Behemoth is a mixture-of-experts model in Meta's Llama 4 line, previewed in April 2025 while still in training, which Meta said it was not yet releasing. Meta describes it as multimodal…

      • Nearly two trillion total parameters, with 288B active per token; Meta presents it as taking Llama to new sizes.
      • Mixture-of-experts model with 16 experts, trained as the teacher model for distilling the smaller Llama 4 models.
      Primary source Announced · no news linked
    2. Released
      Llama 4 · Language, Multimodal · Milestone Llama 4 Maverick

      Llama 4 Maverick is a multimodal language model in Meta's Llama 4 line that takes text and images and returns text, released with open weights in BF16 and FP8 and later offered in a preview of…

      • One of Meta's first Llama models built on mixture of experts: 128 routed experts plus a shared expert in alternating dense and MoE layers, 17B of 400B parameters active.
      • Accepts images alongside text as input; Meta describes native multimodality with early fusion of text and vision tokens.
      Primary source Available · no news linked
    3. Released
      Llama 4 · Language, Multimodal · Milestone Llama 4 Scout

      Llama 4 Scout is a multimodal language model in Meta's Llama 4 line that takes text and images and returns text, released with open weights under the Llama 4 Community License and later offered in a…

      • One of Meta's first Llama models built on a mixture-of-experts architecture, with 16 experts and 17B of its 109B parameters active per token.
      • Accepts images alongside text as input; Meta describes native multimodality with early fusion of text and vision tokens.
      Primary source Available · no news linked

  3. 2024

    4 models
    1. Released
      Llama · Language Llama 3.3

      Llama 3.3 is a text-only multilingual 70B model from Meta, released on 6 December 2024 as open weights under the custom Llama 3.3 Community License, in an instruction-tuned version only. Meta says…

      • Text-only, without the image input of the Llama 3.2 Vision models.
      • Meta says it offers performance similar to Llama 3.1 405B at a fraction of the serving cost.
      Primary source Available · no news linked Some details lack sources
    2. Released
      Llama · Language, Multimodal Llama 3.2

      Llama 3.2 is a set of open-weight Meta models released on 25 September 2024: text-only 1B and 3B models aimed at on-device use, and 11B and 90B Vision models, which Meta describes as the first Llama…

      • Adds image input in the 11B and 90B Vision models; for image plus text tasks English is the only supported language.
      • Adds 1B and 3B text models for mobile and edge devices, enabled for Qualcomm and MediaTek hardware and optimised for Arm processors.
      Primary source Available · no news linked
    3. Released
      Llama · Language Llama 3.1

      Llama 3.1 is a collection of Meta language models in 8B, 70B and 405B sizes, released on 23 July 2024 as open weights under the custom Llama 3.1 Community License and through more than 25 partners…

      • Context length increased from 8K to 128K tokens.
      • Official support for eight languages; the Meta Llama 3 model card named English as the intended language.
      Primary source Available · no news linked
    4. Released
      Llama · Language · Milestone Meta Llama 3

      Meta Llama 3 is the third generation of Meta's Llama language models, released on 18 April 2024 as open weights in 8B and 70B sizes, each pretrained and instruction-tuned, under the custom Llama 3…

      • Uses a new tokenizer with a 128K-token vocabulary, which Meta says encodes text with fewer tokens, and grouped query attention in both sizes.
      • Context length increased from 4K tokens for Llama 2 to 8K tokens.
      Primary source Available · no news linked

  4. 2023

    2 models
    1. Released
      Llama · Language · Milestone Llama 2

      Llama 2 is the second generation of Meta's Llama language models for text, released on 18 July 2023 at 7B, 13B and 70B parameters, each with a dialogue-tuned Llama 2-Chat version. Meta made the…

      • Free for research and commercial use under the Llama 2 Community License, where Llama 1 was offered under a noncommercial research licence.
      • Context length doubled from 2k to 4k tokens.
      Primary source Available · no news linked
    2. Released
      Llama · Language · Milestone LLaMA

      LLaMA is a set of large language models for text from Meta, released on 24 February 2023 in sizes of 7B, 13B, 33B and 65B parameters to help researchers study such models. The weights were shared…

      Primary source Research only · no news linked

  5. 2022

    1 model
    1. Released
      Language · Milestone OPT

      OPT (Open Pre-trained Transformers) is a suite of decoder-only language models from Meta, ranging from 125 million to 175 billion parameters, released on 3 May 2022 together with the metaseq code…

      Primary source Available · no news linked

Lineage

  1. OPT

    No recorded relations.

  2. LLaMA
    Successors
    Llama 2
    Variants and derived
    IDEFICS (Hugging Face) · derived: other
  3. Llama 2
    Predecessors
    LLaMA
    Successors
    Meta Llama 3
    Variants and derived
    PPLX Online LLMs (Perplexity) · derived: fine tune
  4. Meta Llama 3
    Predecessors
    Llama 2
    Successors
    Llama 3.1
  5. Llama 3.1
    Predecessors
    Meta Llama 3
    Successors
    Llama 3.2
    Variants and derived
    Idefics3 (Hugging Face) · derived: other; Llama 3.2 · derived: distillation; Llama-3.1-Nemotron-51B-Instruct (NVIDIA) · derived: distillation; Llama-3.1-Nemotron-70B-Instruct (NVIDIA) · derived: fine tune; Llama-3.1-Nemotron-Nano-8B-v1 (NVIDIA) · derived: fine tune; Llama-3.1-Nemotron-Ultra-253B-v1 (NVIDIA) · derived: distillation
  6. Llama 3.2
    Predecessors
    Llama 3.1
    Successors
    Llama 3.3
    Based on
    Llama 3.1 · derived: distillation
  7. Llama 3.3
    Predecessors
    Llama 3.2
    Variants and derived
    Llama-3.3-Nemotron-Super-49B-v1 (NVIDIA) · derived: distillation; Llama-3.3-Nemotron-Super-49B-v1.5 (NVIDIA) · derived: distillation; Sonar (February 2025) (Perplexity) · derived: fine tune
  8. Llama 4 Behemoth
    Variants and derived
    Llama 4 Maverick · derived: distillation; Llama 4 Scout · derived: distillation
  9. Llama 4 Maverick
    Predecessors
    Llama 3.3
    Successors
    Muse Spark
    Based on
    Llama 4 Behemoth · derived: distillation
  10. Llama 4 Scout
    Predecessors
    Llama 3.3
    Based on
    Llama 4 Behemoth · derived: distillation
  11. Muse Spark
    Predecessors
    Llama 4 Maverick
    Successors
    Muse Spark 1.1
    Variants and derived
    Muse Glimmer · derived: distillation
  12. Muse Spark 1.1
    Predecessors
    Muse Spark
    Successors
    Muse Spark 1.2
  13. Muse Spark 1.2
    Predecessors
    Muse Spark 1.1
    Successors
    Muse Spark 1.3
  14. Muse Glimmer
    Based on
    Muse Spark · derived: distillation
  15. Muse Spark 1.3
    Predecessors
    Muse Spark 1.2

Families

Not in a family

About the organisation

  • Description Meta Platforms, Inc., named Facebook, Inc. until October 2021, runs apps including Facebook, Messenger, Instagram and WhatsApp. Its FAIR research team dates from late 2013. Meta says it has released its Llama foundation models openly and expects to train both open and closed models; Meta Superintelligence Labs builds the Muse series of large language models. Primary source[1] [3] [4] [5] [6]
  • Former name Facebook, Inc. (until ) Primary source[1]Item 5.03 [2]Corporate Information

Company · https://www.meta.com/ · 15 models listed here

Sources

  1. [1] Form 8-K: Current Report (October 28, 2021)

    Other · Meta · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.sec.gov/Archives/edgar/data/1326801/000132680121000071/fb-20211028.htm Current report filed on SEC EDGAR on 2021-10-28; Item 5.03 states that on October 28, 2021 Facebook, Inc. changed its corporate name to Meta Platforms, Inc. Title from the cover page; the HTML title is fb-20211028.

  2. [2] Form 10-K for the fiscal year ended December 31, 2021

    Other · Meta · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.sec.gov/Archives/edgar/data/1326801/000132680122000018/fb-20211231.htm Annual report filed by Meta Platforms, Inc. on SEC EDGAR (filing date 2022-02-03); states that in October 2021 the company changed its corporate name from Facebook, Inc. to Meta Platforms, Inc. Title from the cover page; the HTML title is fb-20211231.

  3. [3] Company info

    Other · Meta

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.meta.com/about/company-info/ Living page; content as seen on 2026-10-01. https://about.meta.com/company-info/ redirects (301) here; requests from the Netherlands are then redirected (302) to the locale version https://www.meta.com/en-gb/about/company-info/, which declares itself canonical.

  4. [4] Celebrating 10 years of FAIR: A decade of advancing the state-of-the-art through open research

    Blog post · Meta · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://ai.meta.com/blog/fair-10-year-anniversary-open-science-meta/ Headline as shown on the page; the HTML title is Ten years of FAIR: Advancing the state-of-the-art through open research. Says the launch of FAIR dates back to late 2013.

  5. [5] Form 10-K for the fiscal year ended December 31, 2025

    Other · Meta · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm Annual report filed by Meta Platforms, Inc. on SEC EDGAR (filing date 2026-01-29, signed 2026-01-28). Title from the cover page; the HTML title is meta-20251231. sec.gov answers 403 to requests without a declared User-Agent, which source health may need to handle.

  6. [6] Introducing Muse Spark: MSL’s First Model, Purpose-Built to Prioritize People

    Announcement · Meta · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/ Meta Newsroom post; title is the headline shown on the page, the HTML title is Introducing Muse Spark: Meta's Most Powerful Model Yet. The page also carries an update dated May 12, 2026 above the original April 8 text.