AI models released in March 2024

14 releases from 13 labs in March 2024, including Grok-1.5, Jamba and Qwen1.5-MoE-A2.7B. Every entry is dated to the lab's official announcement and links to it.

Releases
14
Labs
13
Open weights
9

March 2024

14 releases
  1. xAILLMProprietary

    Grok-1.5

    Grok-1.5 is xAI's update to Grok-1 with improved reasoning and problem solving and long-context understanding of up to 128K tokens, 16 times the previous context length.

    Modalities
    text → text
    Context
    125K context
    At launch
    Consumer app
    • 128,000-token context window
    • Initially for early testers and existing Grok users on X
    • Company timeline lists Grok-1.5 on X on May 15, 2024
    Source: Announcing Grok-1.5 | SpaceXSI
  2. AI21 LabsLLMOpen weights

    Jamba

    The first production-grade model based on Mamba, a hybrid of structured state space and transformer layers with mixture-of-experts, released as an open base model.

    Modalities
    text → text
    Context
    250K context
    Parameters
    MoE: 52B total, 12B active
    License
    Apache-2.0
    At launch
    Open weights
    • 256K-token context window
    • Uses 12B of 52B parameters at inference
    • An instruction-tuned Jamba-Instruct followed on the AI21 platform
    Source: Introducing Jamba: AI21's Groundbreaking SSM-Transformer Model
  3. Alibaba QwenSmall modelOpen weights

    Qwen1.5-MoE-A2.7B

    Qwen1.5-MoE-A2.7B is Qwen's first mixture-of-experts model, with 2.7B activated parameters, created by repurposing Qwen-1.8B. Qwen reports it matches leading 7B models.

    Modalities
    text → text
    Parameters
    MoE: 2.7B active
    License
    Tongyi Qianwen License Agreement
    At launch
    Open weights
    • Contains 2.0B non-embedding parameters, about one-third of Qwen1.5-7B
    • Upcycled from the dense Qwen-1.8B model
    • Weights under the Tongyi Qianwen License Agreement
    Source: Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters
  4. DatabricksLLMOpen weightsRetired

    DBRX

    Open general-purpose language model from Databricks built on a fine-grained mixture-of-experts architecture, released as DBRX Base and DBRX Instruct for download and API use.

    Modalities
    text → text
    Context
    32K context
    Parameters
    MoE: 132B total, 36B active
    License
    Databricks Open Model License
    At launch
    Open weights, API
    Retired
    • Pretrained on 12 trillion tokens of curated data
    • Inference up to twice as fast as LLaMA2-70B
    • Offered to Databricks customers through APIs at launch

    Databricks retired DBRX and DBRX Instruct from pay-per-token serving on April 30, 2025 and from provisioned throughput on December 19, 2025.

    Source: Introducing DBRX: A New State-of-the-Art Open LLM | Databricks Blog

    Retirement: AI models maintenance policy | Databricks on AWS

  5. Stability AICodingOpen weights

    Stable Code Instruct 3B

    Instruction-tuned version of Stable Code 3B that adds natural language interaction for code generation and software development tasks, focusing on Python, JavaScript, Java, C, C++ and Go.

    Modalities
    text → text
    Parameters
    3B
    License
    Stability AI Community License
    At launch
    Open weights
    • Built on Stable Code 3B
    • Weights and code on Hugging Face
    • Commercial use through a Stability AI Membership
    Source: Introducing Stable Code Instruct 3B
  6. SunoMusicProprietary

    Suno v3

    Suno's first model capable of radio-quality music, generating full two-minute songs in seconds, released to all users after testing as v3 Alpha with paid subscribers.

    Modalities
    text → audio
    At launch
    Consumer app
    • Full two-minute songs
    • Tested as v3 Alpha with Pro and Premier members
    • Available to all users at launch
    Source: Introducing v3
  7. xAILLMOpen weights

    Grok-1 (open weights)

    xAI released the base model weights and network architecture of Grok-1, a 314B-parameter mixture-of-experts model trained from scratch, as the raw pretraining checkpoint without fine-tuning for dialogue.

    Modalities
    text → text
    Context
    8K context
    Parameters
    314B (MoE)
    License
    Apache-2.0
    At launch
    Open weights
    • Mixture of 8 experts with 2 used per token (25% of weights active)
    • Pretraining concluded in October 2023 on a custom JAX and Rust stack
    • Weights on Hugging Face (xai-org/grok-1) and code on GitHub (xai-org/grok-1)
    • 8,192-token maximum context per the GitHub README
    Source: Open Release of Grok-1 | SpaceXSI
  8. RecraftImageProprietary

    Recraft V2

    Recraft's image model built for designers, generating both vector and raster images in consistent styles, compared by Recraft against Midjourney v6, DALL-E 3 and open models.

    Modalities
    text → image
    At launch
    Consumer app
    • Generates vector and raster images
    • Designed for brand-consistent graphics
    • Available in the Recraft app
    Source: Recraft V2: First AI Model Built for Designers
  9. CohereLLMOpen weights

    Command R

    Scalable generative model built for retrieval-augmented generation and tool use, with inline citations and a 128K-token context window, plus weights for research use.

    Modalities
    text → text
    Context
    125K context
    Parameters
    35B
    License
    CC-BY-NC-4.0
    At launch
    API, Open weights, Cloud platforms
    • Strong in 10 business languages
    • Designed to pair with Cohere Embed and Rerank
    • Weights on Hugging Face under a non-commercial license
    Source: Command R: RAG at Production Scale
  10. DeepSeekMultimodalOpen weights

    DeepSeek-VL

    DeepSeek-VL is a vision-language model family released in 1.3B and 7B sizes with base and chat variants. The 7B model uses a hybrid SigLIP and SAM vision encoder that supports 1024 by 1024 image input.

    Modalities
    text, image → text
    Parameters
    1.3B, 7B
    License
    DeepSeek Model License
    At launch
    Open weights
    • Four models: 1.3B and 7B, each as base and chat
    • DeepSeek-VL-7B-base is trained on around 400B vision-language tokens
    • Supports commercial use under the DeepSeek Model License
    Source: DeepSeek-VL: Towards Real-World Vision-Language Understanding (GitHub)
  11. Nous ResearchLLMOpen weights

    Hermes-2-Pro-Mistral-7B

    Hermes 2 Pro model from Nous Research built on Mistral 7B, which Nous describes as a master function-calling model, released under Apache 2.0.

    Modalities
    text → text
    Parameters
    7B
    License
    Apache-2.0
    At launch
    Open weights
    • Focused on function calling
    • Built on Mistral 7B
    • Later followed by Llama 3-based Hermes 2 Pro models
    Source: Releases | Nous Research
  12. Inflection AILLMProprietary

    Inflection-2.5

    Inflection-2.5 is Inflection AI's upgraded in-house model for its Pi assistant, rolled out to all Pi users with improved coding and mathematics over Inflection-1 and launched alongside real-time web search in Pi.

    Modalities
    text → text
    At launch
    Consumer app
    • Available to all Pi users at pi.ai, on iOS, on Android and in a new desktop app
    • Pi added real-time web search alongside the upgrade
    • On March 19, 2024, Inflection said it would host Inflection-2.5 on Microsoft Azure for API access
    Source: Inflection-2.5: meet the world's best personal AI
  13. 01.AILLMOpen weights

    Yi-9B

    Yi-9B is an open bilingual base model from 01.AI, continuously trained from Yi-6B on 0.8T tokens and positioned as the Yi series model strongest at coding and math.

    Modalities
    text → text
    Context
    4K context
    Parameters
    9B
    License
    Apache-2.0
    At launch
    Open weights
    • Default context window of 4K tokens
    • Yi-9B-200K long-context variant was open-sourced on March 16, 2024
    • Weights on Hugging Face, ModelScope, and wisemodel
    Source: GitHub - 01-ai/Yi: A series of large language models trained from scratch by developers @01-ai
  14. AnthropicLLMProprietary

    Claude 3 Opus, Sonnet, and Haiku

    The Claude 3 family comprises Claude 3 Haiku, Sonnet, and Opus in ascending order of capability. It was the first Anthropic generation with vision input, and Anthropic reported fewer unnecessary refusals and improved accuracy over Claude 2.1.

    Modalities
    text, image → text
    Context
    200K context
    At launch
    API, Consumer app, Cloud platforms
    • 200K-token context window at launch; inputs over 1M tokens offered to select customers
    • Opus and Sonnet available at launch on claude.ai and the Claude API; Haiku followed on March 13, 2024
    • Sonnet on Amazon Bedrock at launch and in private preview on Google Cloud Vertex AI
    • Claude API became generally available in 159 countries
    Source: Introducing the next generation of Claude

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 March 2024." The AI Engineering, updated Mar 28, 2024. https://theaiengineering.com/releases/2024/03

HTML link

<a href="https://theaiengineering.com/releases/2024/03">AI models released in March 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