AI models released in October 2024
20 releases from 15 labs in October 2024, including Recraft V3, Stable Diffusion 3.5 Medium and Aya Expanse. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 20
- Labs
- 15
- Open weights
- 11
October 2024
20 releasesRecraft V3
Image model trained from scratch with major gains in text generation, able to render long text in images, plus anatomical accuracy, prompt understanding and vector image generation.
- Modalities
- text → image
- At launch
- Consumer app, API
- Generates long texts, not just a few words, within images
- Supports vector art generation and style consistency
- Launched with a Recraft API for developers
Stable Diffusion 3.5 Medium
2.5B-parameter Stable Diffusion 3.5 model with an improved MMDiT-X architecture, designed to run on standard consumer hardware and generate images between 0.25 and 2 megapixels.
- Modalities
- text → image
- Parameters
- 2.5B
- License
- Stability AI Community License
- At launch
- Open weights, API
- 2.5B parameters
- Requires about 9.9 GB of VRAM excluding text encoders
- Released one week after SD 3.5 Large
Aya Expanse
Family of multilingual open-weights models from Cohere For AI covering 23 languages, built by combining data arbitrage, multilingual preference training and model merging.
- Modalities
- text → text
- Parameters
- 8B and 32B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Released on Kaggle and Hugging Face
- Made available to chat with on WhatsApp in January 2025
- Weights published under a non-commercial license
GLM-4-Voice
GLM-4-Voice is Z.ai's end-to-end voice model that understands and generates Chinese and English speech for real-time voice conversation. Users can change emotion, intonation, speech rate, and dialect by instruction.
- Modalities
- text, audio → text, audio
- Parameters
- 9B
- License
- GLM-4-Voice License
- At launch
- Open weights
- Speech tokenizer converts each second of audio into 12.5 discrete tokens
- GLM-4-Voice-9B is pre-trained on speech based on GLM-4-9B
- Streaming decoder can start generating after as few as 10 audio tokens
Llama 3.2 1B and 3B quantized
Meta released quantized versions of Llama 3.2 1B and 3B, its first lightweight quantized Llama models, built with QAT plus LoRA and with SpinQuant to run on mobile devices.
- Modalities
- text → text
- Parameters
- 1B, 3B
- License
- Llama 3.2 Community License
- At launch
- Open weights
- 2 to 4 times faster than the original BF16 models on average
- Average 56% smaller model size and 41% lower memory use
- Model card lists an 8K context length for the quantized models
Act-One
Tool inside Gen-3 Alpha for generating expressive character performances from a single driving video of an actor and a character image, without motion capture or rigging.
- Modalities
- video, image → video
- At launch
- Consumer app
- Captures eye-lines, micro expressions, pacing and delivery
- Rolled out gradually to users from October 2024
- Succeeded by Act-Two in July 2025
Claude 3.5 Sonnet (upgraded) and Claude 3.5 Haiku
Anthropic released an upgraded Claude 3.5 Sonnet with gains in agentic coding and tool use, plus the new Claude 3.5 Haiku. Claude 3.5 Sonnet was the first Anthropic model to offer computer use, in public beta.
- Modalities
- text, image → text
- At launch
- API, Consumer app, Cloud platforms
- API model IDs claude-3-5-sonnet-20241022 and claude-3-5-haiku-20241022
- Computer use public beta on the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI
- Claude 3.5 Haiku reached the Claude API on November 4, 2024 as a text-only model
- Upgraded Claude 3.5 Sonnet available to all Claude users at launch
Embed 3 (multimodal)
Update that makes Cohere's Embed 3 models multimodal, generating embeddings from both text and images in the same vector space for enterprise search.
- Modalities
- text, image → embedding
- At launch
- API
- Text and image embeddings share one vector space
- Targets charts, product catalogs and design files
- Later offered on Amazon Bedrock
Mochi 1 (preview)
Research preview of Genmo's open text-to-video model, a 10B-parameter diffusion model built on an Asymmetric Diffusion Transformer architecture with strong motion quality and prompt adherence.
- Modalities
- text → video
- Parameters
- 10B
- License
- Apache-2.0
- At launch
- Open weights, Consumer app
- 10B parameters, trained from scratch
- Generates 480p video at 30 fps for up to 5.4 seconds
- Hosted playground launched alongside the weights
Stable Diffusion 3.5 Large and Large Turbo
Open Stable Diffusion 3.5 release led by the 8.1B-parameter Large base model for professional 1MP work and a distilled Large Turbo variant that generates images in four steps.
- Modalities
- text → image
- Parameters
- 8.1B
- License
- Stability AI Community License
- At launch
- Open weights, API, Cloud platforms
- Stable Diffusion 3.5 Large has 8.1B parameters
- Large Turbo generates in 4 steps
- Free commercial use for smaller organizations under the Stability AI Community License
Granite 3.0
Third generation of IBM's open language models, released as dense 2B and 8B models and sparse mixture-of-experts 1B and 3B models in base and instruct versions.
- Modalities
- text → text
- Context
- 4K context
- Parameters
- 2B and 8B dense; MoE 1B-A400M and 3B-A800M
- License
- Apache-2.0
- At launch
- Open weights
- 4,096-token sequence length
- MoE models activate 400M and 800M parameters
- Released under Apache 2.0
Granite Guardian 3.0
Safety models fine-tuned from Granite 3.0 Instruct to detect risks in prompts and responses, including harm, jailbreaking and RAG issues such as groundedness and context relevance.
- Modalities
- text → text
- Parameters
- 2B and 8B
- License
- Apache-2.0
- At launch
- Open weights
- Released in 2B and 8B sizes
- Covers context relevance, groundedness and answer relevance for RAG
- Released under Apache 2.0
Janus
Janus is an autoregressive framework that unifies multimodal understanding and image generation by decoupling visual encoding into separate pathways while using a single transformer. DeepSeek released it as Janus-1.3B.
- Modalities
- text, image → text, image
- Parameters
- 1.3B
- License
- DeepSeek Model License
- At launch
- Open weights
- Separate visual encoding paths for understanding and generation
- Single unified transformer architecture
- Code under the MIT License; model use under the DeepSeek Model License
Ministral 3B
Compact 3B edge model introduced as one of les Ministraux for local, privacy-first and low-latency applications such as on-device translation and offline assistants.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 3B
- At launch
- API
- Supports up to 128K tokens of context
- Offered through the API rather than as open weights
- API model ID ministral-3b-2410
Ministral 8B
Edge-focused 8B model introduced with Ministral 3B for on-device and low-latency use, featuring an interleaved sliding-window attention pattern for memory-efficient inference.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 8B
- License
- Mistral Research License
- At launch
- Open weights, API
- Supports up to 128K tokens of context
- Suited to function calling in multi-step agent workflows
- API model ID ministral-8b-2410
Inflection 3.0 (Pi 3.0 and Productivity 3.0)
Inflection AI's enterprise-grade model generation, launched with Inflection for Enterprise and a commercial API, comprising Pi 3.0 for emotionally intelligent conversation and Productivity 3.0 for instruction following and JSON output.
- Modalities
- text → text
- At launch
- API, Consumer app, Enterprise
- API config IDs inflection_3_pi and inflection_3_productivity at developers.inflection.ai
- Both models supported an 8k context window, per the developer docs
- Pi 3.0 was described as the model powering the Pi experience
- Runs on Intel Gaudi 3 accelerators and Intel Tiber AI Cloud
Reka Flash (October 2024 update)
An updated Reka Flash with function calling, structured output, stronger instruction following, native audio understanding in video, and support for 3 to 5 minute videos, up from about one minute in the first version.
- Modalities
- text, image, video, audio → text
- Parameters
- 21B
- At launch
- API, Consumer app, Enterprise
- Interleaved text, image, video and audio input with a 128K-token context length
- Available on Reka Chat and the Reka API, plus on-premises and private cloud deployment
- Speech output was experimental and not in the public API at launch
FLUX1.1 [pro]
Successor to FLUX.1 [pro] with faster generation and improved image quality, prompt adherence and diversity, launched together with general availability of the beta BFL API.
- Modalities
- text → image
- At launch
- API, Cloud platforms
- Tested under the codename blueberry before launch
- Released with the beta BFL API
- Also offered through Together.ai, Replicate, fal.ai and Freepik
gpt-4o-realtime-preview and gpt-4o-audio-preview
OpenAI launched the Realtime API in public beta with gpt-4o-realtime-preview, a GPT-4o model for low-latency speech-to-speech over WebSockets, and announced gpt-4o-audio-preview to bring audio input and output to the Chat Completions API.
- Modalities
- text, audio → text, audio
- At launch
- API
- Realtime API public beta opened to all paid developers, with six preset voices and function calling
- gpt-4o-audio-preview released in Chat Completions on Oct 17, 2024 (API changelog)
- Five more voices added Oct 30, 2024; updated 2024-12-17 snapshots followed
- OpenAI announced Realtime API general availability on Aug 28, 2025
Whisper large-v3-turbo
Whisper large-v3-turbo is an open-weights speech recognition model that fine-tunes a pruned Whisper large-v3, cutting decoder layers from 32 to 4 for faster transcription with minimal accuracy degradation.
- Modalities
- audio → text
- Parameters
- 809M
- License
- MIT
- At launch
- Open weights
- Called turbo in the whisper package; weights on Hugging Face as openai/whisper-large-v3-turbo
- 809M parameters versus 1550M for large-v3
- Not trained for translation tasks; it returns the original language
- Code and weights shipped in the whisper v20240930 GitHub release
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 October 2024." The AI Engineering, updated Oct 30, 2024. https://theaiengineering.com/releases/2024/10
HTML link
<a href="https://theaiengineering.com/releases/2024/10">AI models released in October 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