AI models released in September 2024

23 releases from 12 labs in September 2024, including LFM-1B, LFM-3B, and LFM-40B, SAM 2.1 and CogView3-Plus-3B. Every entry is dated to the lab's official announcement and links to it.

Releases
23
Labs
12
Open weights
19

September 2024

23 releases
  1. Liquid AILLMProprietary

    LFM-1B, LFM-3B, and LFM-40B

    Liquid AI's first series of Liquid Foundation Models, non-transformer language models built from dynamical systems and signal processing principles, released as dense 1.3B and 3.1B models plus a 40.3B mixture-of-experts model.

    Modalities
    text → text
    Parameters
    1B: 1.3B; 3B: 3.1B; 40B: 40.3B MoE (12B active)
    At launch
    Consumer app, Enterprise
    • Preview release optimized for a 32k-token context length
    • LFM-40B activates 12B parameters per token
    • Available on Liquid Playground; weights not open-sourced
    • Liquid AI now marks this generation as deprecated
    Source: Liquid Foundation Models: Our First Series of Generative AI Models
  2. MetaMultimodalOpen weights

    SAM 2.1

    SAM 2.1 is an updated set of Segment Anything Model 2 checkpoints with stronger performance, released with the SAM 2 training code and web demo code as the SAM 2.1 Developer Suite.

    Modalities
    image, video → image, video
    License
    Apache-2.0
    At launch
    Open weights
    • New checkpoints require the latest SAM 2 repository code
    • Training and fine-tuning code released
    • Front-end and back-end code for the SAM 2 web demo released
    Source: facebookresearch/sam2 GitHub repository
  3. Z.aiImageOpen weights

    CogView3-Plus-3B

    CogView-3Plus-3B is Z.ai's open-weights text-to-image model, released together with the CogView3 text-to-image system. Z.ai added a diffusers version of the model on October 13, 2024.

    Modalities
    text → image
    Parameters
    3B
    License
    Apache-2.0
    At launch
    Open weights
    • Released alongside CogView3 on September 29, 2024
    • Diffusers support added October 13, 2024
    • Weights under the Apache 2.0 license
    Source: CogView4 & CogView3 & CogView-3Plus (GitHub)
  4. OpenAIMultimodalProprietary

    omni-moderation-latest

    omni-moderation-latest is a GPT-4o-based model for OpenAI's free Moderation API that classifies both text and images, adds two text-only harm categories, and improves accuracy, especially for non-English content.

    Modalities
    text, image → text
    At launch
    API
    • Snapshot ID omni-moderation-2024-09-26
    • Image inputs supported for violence, self-harm, and sexual categories
    • Adds illicit and illicit/violent text categories
    • Free for all developers, with rate limits by usage tier
    Source: Upgrading the Moderation API with our new multimodal moderation model
  5. MetaMultimodalOpen weights

    Llama 3.2 11B and 90B Vision

    Llama 3.2 Vision adds image understanding to Llama with 11B and 90B parameter models that take text and images as input and produce text output.

    Modalities
    text, image → text
    Context
    125K context
    Parameters
    11B, 90B
    License
    Llama 3.2 Community License
    At launch
    Open weights, Cloud platforms
    • Model card lists a 128K context length
    • English is the only supported language for image plus text tasks
    • Released with the first official Llama Stack distributions
    Source: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
  6. MetaSmall modelOpen weights

    Llama 3.2 1B and 3B

    Llama 3.2 1B and 3B are lightweight text-only models designed for edge and mobile devices, supporting a 128K token context window.

    Modalities
    text → text
    Context
    125K context
    Parameters
    1B, 3B
    License
    Llama 3.2 Community License
    At launch
    Open weights, Cloud platforms
    • Officially supports eight languages
    • Pretraining data cutoff of December 2023
    • Released alongside the Llama 3.2 Vision models
    Source: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
  7. MetaMultimodalOpen weights

    Llama Guard 3 11B Vision

    Llama Guard 3 11B Vision is a safety classifier fine-tuned from Llama 3.2 11B that filters text plus image prompts and text responses, supporting Llama 3.2's image understanding.

    Modalities
    text, image → text
    Parameters
    11B
    License
    Llama 3.2 Community License
    At launch
    Open weights
    • Fine-tuned from the Llama 3.2 11B pretrained model
    • Predicts safety labels across 13 MLCommons-based hazard categories
    Source: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
  8. MetaSmall modelOpen weights

    Llama Guard 3 1B

    Llama Guard 3 1B is a compact content safety classifier based on Llama 3.2 1B, with a pruned and quantized version for deployment on mobile devices.

    Modalities
    text → text
    Parameters
    1B
    License
    Llama 3.2 Community License
    At launch
    Open weights
    • Pruned and quantized version shrinks from 2,858 MB to 438 MB
    • Covers 13 hazard categories in eight languages
    Source: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
  9. Ai2MultimodalOpen weights

    Molmo

    Family of open vision-language models trained without distillation from proprietary VLMs, notable for learning to point at what they perceive in images.

    Modalities
    text, image → text
    Parameters
    1B (MolmoE), 7B and 72B
    License
    Apache-2.0
    At launch
    Open weights
    • Weights, captioning data and fine-tuning data released
    • Released as MolmoE-1B, Molmo-7B-O, Molmo-7B-D and Molmo-72B
    • Built from pretrained vision encoders and language-only LLMs
    Source: Molmo | Ai2
  10. GoogleLLMProprietary

    Gemini 1.5 Pro-002 and 1.5 Flash-002

    Google released updated production versions of Gemini 1.5 Pro and 1.5 Flash that built on its experimental releases, with improved overall quality, larger gains in math, long context and vision, and lower latency.

    Modalities
    text, image, audio, video → text
    At launch
    API, Cloud platforms
    • Model IDs gemini-1.5-pro-002 and gemini-1.5-flash-002
    • The gemini-1.5-pro-latest and gemini-1.5-flash-latest aliases were updated to point to the -002 models
    • Improved the models' ability to follow user instructions while balancing safety
    Source: Updated production-ready Gemini models, reduced 1.5 Pro pricing, increased rate limits, and more
  11. NVIDIALLMOpen weights

    Llama-3.1-Nemotron-51B-Instruct

    51B model derived from Meta's Llama-3.1-70B with neural architecture search and knowledge distillation, built to run efficiently on a single H100 GPU.

    Modalities
    text → text
    Parameters
    51B
    License
    NVIDIA Open Model License
    At launch
    Open weights, API
    • Uses block distillation and knowledge distillation
    • Offered as an NVIDIA NIM microservice
    • Weights on Hugging Face
    Source: Advancing the Accuracy-Efficiency Frontier with Llama-3.1-Nemotron-51B | NVIDIA Technical Blog
  12. Alibaba QwenLLMOpen weights

    Qwen2.5

    Qwen2.5 is Qwen's family of dense decoder-only language models from 0.5B to 72B parameters, pre-trained on up to 18T tokens. It added 3B, 14B, and 32B sizes and launched alongside Qwen2.5-Coder and Qwen2.5-Math.

    Modalities
    text → text
    Context
    128K context
    Parameters
    0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B
    License
    Apache-2.0 (3B and 72B use the Qwen Research License and Qwen License)
    At launch
    API, Open weights
    • Full 131,072-token context with 8,192-token generation for the larger models
    • All open models except 3B and 72B under Apache 2.0
    • Qwen2-VL-72B weights were also open-sourced the same day
    Source: Qwen2.5: A Party of Foundation Models!
  13. Alibaba QwenCodingOpen weights

    Qwen2.5-Coder

    Qwen2.5-Coder is the next generation of Qwen's open code models, renaming CodeQwen to Qwen-Coder. It continues Qwen2.5 training on 5.5T tokens of code, text-code grounding, and synthetic data.

    Modalities
    text → text
    Context
    125K context
    Parameters
    1.5B, 7B
    License
    Apache-2.0
    At launch
    Open weights
    • Supports up to 128K tokens of context
    • Covers 92 programming languages
    • Released in 1.5B and 7B sizes, with 32B announced as coming soon
    Source: Qwen2.5-Coder: Code More, Learn More!
  14. Alibaba QwenReasoningOpen weights

    Qwen2.5-Math

    Qwen2.5-Math is Qwen's upgraded math model series with base and instruct models at 1.5B, 7B, and 72B plus the Qwen2.5-Math-RM-72B reward model. It solves English and Chinese math problems with chain-of-thought and tool-integrated reasoning.

    Modalities
    text → text
    Context
    4K context
    Parameters
    1.5B, 7B, 72B
    License
    Apache-2.0 (72B uses the Qwen License Agreement)
    At launch
    Open weights
    • Supports both chain-of-thought and tool-integrated reasoning
    • Pre-trained on Qwen Math Corpus v2 with a 4K context length
    • Qwen does not recommend it for non-math tasks
    Source: Qwen2.5-Math: The world's leading open-sourced mathematical LLMs
  15. Mistral AILLMOpen weights

    Mistral Small 24.09

    Upgrade of Mistral Small v24.02 released with open weights under the Mistral Research License as part of a September 2024 update to la Plateforme and le Chat.

    Modalities
    text → text
    Context
    32K context
    Parameters
    22B
    License
    Mistral Research License
    At launch
    Open weights, API
    • Self-deployment permitted for non-commercial use
    • API model ID mistral-small-2409
    • Released alongside Pixtral 12B access in le Chat
    Source: AI in abundance | Mistral AI
  16. NVIDIAMultimodalOpen weights

    NVLM 1.0 (NVLM-D-72B)

    Family of frontier-class multimodal language models from NVIDIA's applied deep learning research team, with the 72B NVLM-D model's weights and training code open-sourced.

    Modalities
    text, image → text
    Parameters
    72B
    License
    CC-BY-NC-4.0
    At launch
    Open weights
    • Handles vision-language tasks while reporting text-only results
    • Training code released in Megatron-Core
    • Weights on Hugging Face under CC BY-NC 4.0
    Source: NVLM: Open Frontier-Class Multimodal LLMs - NVIDIA ADLR
  17. Mistral AIMultimodalOpen weights

    Pixtral 12B

    Mistral AI's first multimodal model, a 12B decoder based on Mistral NeMo paired with a new vision encoder that accepts images of variable size and aspect ratio.

    Modalities
    text, image → text
    Context
    125K context
    Parameters
    12B decoder + 400M vision encoder
    License
    Apache-2.0
    At launch
    Open weights, API, Consumer app
    • Weights first published September 11, 2024, per Mistral's model card
    • Supports multiple images within a 128K-token context
    • Made available in le Chat
    • API model ID pixtral-12b-2409
    Source: Pixtral 12B | Mistral AI
  18. GoogleLLMOpen weights

    DataGemma

    DataGemma is a pair of Gemma 2-based open models that ground statistical answers in Google's Data Commons knowledge graph to reduce hallucinations. One variant uses retrieval-interleaved generation (RIG), the other retrieval-augmented generation (RAG).

    Modalities
    text → text
    Parameters
    27B
    License
    Gemma Terms of Use
    At launch
    Open weights
    • Data Commons holds over 240 billion data points from sources such as the UN, WHO and CDC
    • RIG model annotates statistics with natural language queries to Data Commons
    • Released on Hugging Face for academic and research use
    Source: DataGemma: Using real-world data to address AI hallucinations
  19. OpenAIReasoningProprietary

    OpenAI o1-preview and o1-mini

    o1-preview and o1-mini are the first models in OpenAI's o1 series, trained with reinforcement learning to reason through problems before responding. o1-mini is a faster version that OpenAI describes as particularly effective at coding.

    Modalities
    text → text
    Context
    125K context
    At launch
    Consumer app, API
    • API model IDs o1-preview-2024-09-12 and o1-mini-2024-09-12, first for usage tier 5 developers
    • Available to ChatGPT Plus and Team at launch; Enterprise and Edu the following week
    • API launched without function calling, streaming, or system messages
    • Both models have since been retired from the API
    Source: Introducing OpenAI o1
  20. SalesforceLLMOpen weights

    xLAM

    Family of large action models optimized for function calling and AI agents, released for research in sizes for mobile use through high-performance deployments.

    Modalities
    text → text
    Parameters
    7B, 8x7B and 8x22B
    License
    CC-BY-NC-4.0
    At launch
    Open weights
    • Variants include xLAM-7b-r, xLAM-8x7b-r and xLAM-8x22b-r
    • Followed the July 17, 2024 release of the xLAM-1b-fc-r and xLAM-7b-fc-r function-calling models
    • Released for research under CC BY-NC 4.0
    Source: xLAM: A Family of Large Action Models for AI Agents - Salesforce
  21. DeepSeekLLMOpen weights

    DeepSeek-V2.5

    DeepSeek-V2.5 merges DeepSeek-V2-Chat-0628 and DeepSeek-Coder-V2-0724 into one model that combines general conversation and coding abilities. It is better aligned with human preferences and improves writing and instruction following.

    Modalities
    text → text
    License
    DeepSeek Model License
    At launch
    API, Consumer app, Open weights
    • API users can reach it through either deepseek-chat or deepseek-coder
    • Strengthened resistance to jailbreak attacks while reducing over-refusal
    • Weights on Hugging Face
    Source: DeepSeek-V2.5: A New Open-Source Model Combining General and Coding Capabilities
  22. 01.AICodingOpen weights

    Yi-Coder

    Yi-Coder is 01.AI's open-source code model series in 1.5B and 9B sizes, released as base and chat versions with long-context support for code understanding and generation.

    Modalities
    text → text
    Context
    128K context
    Parameters
    1.5B, 9B
    License
    Apache-2.0
    At launch
    Open weights
    • Maximum context length of 128K tokens
    • Supports 52 major programming languages
    • Base and chat versions for both sizes on Hugging Face, ModelScope, and wisemodel
    Source: GitHub - 01-ai/Yi-Coder: Yi-Coder is a series of open-source code language models that delivers state-of-the-art coding performance with fewer than 10 billion parameters.
  23. Ai2LLMOpen weights

    OLMoE

    First mixture-of-experts model in the OLMo family, developed with Contextual AI and released with open data, code, logs and intermediate checkpoints.

    Modalities
    text → text
    Parameters
    MoE: 7B total, 1B active
    License
    Apache-2.0
    At launch
    Open weights
    • Trained on 5 trillion tokens
    • 244 pretraining checkpoints released, one every 5,000 steps
    • Instruct version previewed the Tülu 3 post-training pipeline
    Source: OLMoE: An open, small, and state-of-the-art mixture-of-experts model | Ai2

Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.

Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

Citation

The AI Engineering Editors. "AI models released in September 2024." The AI Engineering, updated Sep 30, 2024. https://theaiengineering.com/releases/2024/09

HTML link

<a href="https://theaiengineering.com/releases/2024/09">AI models released in September 2024</a> (The AI Engineering)
Full dataset
1,167 releases from 55 labs
Coverage
to
Updated
License
CC BY 4.0
Publisher
The AI Engineering

Archive