AI models released in December 2024
29 releases from 14 labs in December 2024, including DeepSeek-V3, QVQ-72B-Preview and Gemini 2.0 Flash Thinking. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 29
- Labs
- 14
- Open weights
- 13
December 2024
29 releasesDeepSeek-V3
DeepSeek-V3 is an open-weights mixture-of-experts model built on Multi-head Latent Attention and the DeepSeekMoE architecture, pre-trained on 14.8T tokens. It replaced the model behind the deepseek-chat API and ran about 3 times faster than V2.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 671B total, 37B active
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- 671B total parameters with 37B activated per token
- Generates about 60 tokens per second, per DeepSeek
- Code under the MIT License; weights under the DeepSeek Model License, which permits commercial use
QVQ-72B-Preview
QVQ-72B-Preview is an experimental open-weights model for multimodal reasoning, built on Qwen2-VL-72B. It improves math-related visual reasoning compared with Qwen2-VL-72B-Instruct.
- Modalities
- text, image → text
- Parameters
- 72B
- License
- Qwen License Agreement
- At launch
- Open weights
- Experimental research model with known limitations such as language mixing
- Built upon Qwen2-VL-72B
- Weights under the Qwen License Agreement
Gemini 2.0 Flash Thinking
Gemini 2.0 Flash Thinking was an experimental test-time compute model that showed its thought process while generating a response, producing stronger reasoning than the base Gemini 2.0 Flash.
- Modalities
- text → text
- At launch
- API, Consumer app
- Released for public preview in the Gemini API and Google AI Studio
- Updated as gemini-2.0-flash-thinking-exp-01-21 on Jan 21, 2025
- Added to the Gemini app model dropdown in February 2025
Eleven Flash v2 and v2.5
Ultra-low-latency text-to-speech models that generate speech in 75ms plus application and network latency, recommended by ElevenLabs for conversational voice agents and real-time use.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Flash v2 is English only; Flash v2.5 supports 32 languages
- Model IDs eleven_flash_v2 and eleven_flash_v2_5
- Slightly lower quality than Turbo with lower latency
Granite 3.1
Update of the Granite 3 language models that extends every size to a 128K-token context window, released in base and instruct versions.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 2B and 8B dense; MoE 1B-A400M and 3B-A800M
- License
- Apache-2.0
- At launch
- Open weights
- 128K-token context across dense and MoE models
- Tuned with synthetic data for long-context problems
- Released under Apache 2.0
Granite Embedding
IBM's first Granite embedding models for retrieval, released as two English and two multilingual dense bi-encoder models under Apache 2.0.
- Modalities
- text → embedding
- Parameters
- 30M and 125M English; 107M and 278M multilingual
- License
- Apache-2.0
- At launch
- Open weights
- English models in 30M and 125M sizes
- Multilingual models in 107M and 278M sizes
- Followed by sparse and R2 versions in 2025
Granite Guardian 3.1
Updated Granite Guardian safety models built on Granite 3.1 for judging risks in prompts and responses, including harm and RAG hallucination checks.
- Modalities
- text → text
- Parameters
- 2B and 8B
- License
- Apache-2.0
- At launch
- Open weights
- Released in 2B and 8B sizes
- Detects risks in prompts, responses and RAG outputs
- Released under Apache 2.0
Phoenix 1.0
Full release of Leonardo.Ai's Phoenix foundational image model, with improved aesthetics over the preview, launched on the platform alongside Flux Dev and Flux Schnell.
- Modalities
- text → image
- At launch
- Consumer app, API
- Offered in the API as phoenix-v1.0 alongside phoenix-v0.9
- Released on Replicate with Motion 2.0 on June 10, 2025
gpt-4o-mini-realtime-preview and gpt-4o-mini-audio-preview
OpenAI brought GPT-4o mini to the Realtime API beta and to audio in Chat Completions as preview speech models, alongside a new gpt-4o-realtime-preview-2024-12-17 snapshot and WebRTC support for the Realtime API.
- Modalities
- text, audio → text, audio
- At launch
- API
- Realtime API model ID gpt-4o-mini-realtime-preview-2024-12-17
- Chat Completions model ID gpt-4o-mini-audio-preview-2024-12-17
- Announced with o1 in the API, Preference Fine-Tuning, and WebRTC support for the Realtime API
- Both snapshots are now listed as shut down on their API docs model pages
Veo 2
Veo 2 is Google DeepMind's second-generation video generation model, producing clips at resolutions up to 4K that can extend to minutes in length. It rolled out first through Google Labs' VideoFX.
- Modalities
- text, image → video
- At launch
- Consumer app, API, Cloud platforms
- Supports resolutions up to 4K
- Initial rollout in VideoFX through a waitlist
- Generally available in the Gemini API as veo-2.0-generate-001 from April 9, 2025
Command R7B
The smallest, fastest and final model in Cohere's R series, built for RAG with citations, tool use and agents on modest hardware.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 7B
- License
- CC-BY-NC-4.0
- At launch
- API, Open weights
- 128K-token context length
- Focus on math, reasoning, code and multilingual tasks
- Research weights released as c4ai-command-r7b-12-2024
DeepSeek-VL2
DeepSeek-VL2 is a series of mixture-of-experts vision-language models that improves on DeepSeek-VL in visual question answering, OCR, document, table, and chart understanding, and visual grounding. It comes in Tiny, Small, and standard variants.
- Modalities
- text, image → text
- Parameters
- MoE: 1.0B, 2.8B, 4.5B active (Tiny, Small, VL2)
- License
- DeepSeek Model License
- At launch
- Open weights
- Three variants with 1.0B, 2.8B, and 4.5B activated parameters
- Uses a dynamic tiling strategy for high-resolution images
- Model use is subject to the DeepSeek Model License, which supports commercial use
Grok-2 (grok-2-1212) and Grok-2 Vision (grok-2-vision-1212)
A new version of Grok-2, three times faster with improved accuracy, instruction following, and multilingual ability, rolled out free to all X users and was added to the API as grok-2-1212 and grok-2-vision-1212.
- Modalities
- text, image → text
- At launch
- Consumer app, API
- API model IDs: grok-2-1212 and grok-2-vision-1212
- Free for all users on X, with higher limits for Premium and Premium+
- Docs release notes list API availability on December 14, 2024
Phi-4
Phi-4 is a 14B Microsoft small language model specializing in complex reasoning, particularly math, with gains attributed to synthetic training data, curated organic data and improved post-training methods.
- Modalities
- text → text
- Context
- 16K context
- Parameters
- 14B
- License
- MIT
- At launch
- Open weights, API
- 14B parameter dense decoder-only model with a 16K token context length
- Announcement updated January 8, 2025, listing availability on Azure AI Foundry and Hugging Face
- Weights released under the MIT license
Gemini 2.0 Flash
Gemini 2.0 Flash was the first model in the Gemini 2.0 family, a low-latency workhorse built for the agentic era with native tool use, released first as an experimental model alongside a new Multimodal Live API.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms, Consumer app
- Experimental release in the Gemini API, Google AI Studio and Vertex AI, plus a chat-optimized version in the Gemini app
- Native image generation and text-to-speech output were initially limited to early-access partners
- Generally available in the Gemini API and Vertex AI on Feb 5, 2025 with a 1 million token context window
- Shut down in the Gemini API on June 1, 2026
DeepSeek-V2.5-1210
DeepSeek-V2.5-1210 is the final update in the DeepSeek-V2.5 series, improving math, coding, writing, and roleplay. DeepSeek released it alongside Internet Search on its web chat and described it as the close of the V2 line.
- Modalities
- text → text
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- Upgraded the deepseek-chat API model
- Optimized file upload and webpage summarization experiences
- Weights on Hugging Face as DeepSeek-V2.5-1210
Aurora
Aurora is xAI's image generation model for Grok, an autoregressive mixture-of-experts network trained to predict the next token from interleaved text and image data, with support for editing user-provided images.
- Modalities
- text, image → image
- At launch
- Consumer app
- Launched on X in select countries, rolling out to all users within a week
- Trained on billions of internet examples, with a focus on photorealistic rendering
- Docs release notes list the image generation model on the API on March 19, 2025
Sora (Sora Turbo)
Sora Turbo is a faster version of OpenAI's text-to-video model, released as a standalone product at sora.com for ChatGPT Plus and Pro users. It generates video from text, images, or existing video, moving Sora out of research preview.
- Modalities
- text, image, video → video
- At launch
- Consumer app
- Generates videos up to 1080p and up to 20 seconds in widescreen, vertical, or square formats
- Significantly faster than the Sora model OpenAI previewed in February 2024
- Includes a storyboard tool; all generated videos carry C2PA metadata
- OpenAI notes the Sora product is no longer available as of April 26, 2026
Llama 3.3 70B
Llama 3.3 is a 70B instruction-tuned multilingual text model with a 128K context window, released as a single model in the Llama 3 line.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 70B
- License
- Llama 3.3 Community License
- At launch
- Open weights, Cloud platforms
- Officially supports eight languages, including English, German, Hindi and Thai
- Pretrained on about 15 trillion tokens with a December 2023 cutoff
- Text-only input; outputs text and code
OpenAI o1 (with o1 pro mode)
OpenAI o1 is the next model in OpenAI's o1 reasoning series, succeeding o1-preview in ChatGPT. It launched alongside the ChatGPT Pro plan, which added o1 pro mode, a version of o1 that uses more compute to think harder.
- Modalities
- text, image → text
- Context
- 200K context
- At launch
- Consumer app
- Reached the API on Dec 17, 2024 as o1-2024-12-17, rolling out to usage tier 5 developers
- API version supports function calling, Structured Outputs, developer messages, vision, and reasoning_effort
- 200,000-token context window per the o1 API docs model page
- o1 pro mode is available to ChatGPT Pro subscribers
PaliGemma 2
PaliGemma 2 is Google's second-generation open vision-language model, built on Gemma 2 and offered in three sizes and three input resolutions. It adds detailed long captioning and is designed for easy fine-tuning.
- Modalities
- text, image → text
- Parameters
- 3B, 10B, 28B
- License
- Gemma Terms of Use
- At launch
- Open weights
- Input resolutions of 224px, 448px and 896px
- Drop-in replacement for existing PaliGemma users
- Pre-trained models on Hugging Face and Kaggle
Arctic Embed 2.0
Multilingual second generation of Snowflake's Arctic Embed retrieval models, released as arctic-embed-l-v2.0 and arctic-embed-m-v2.0 for enterprise-grade multilingual search and retrieval.
- Modalities
- text → embedding
- License
- Apache-2.0
- At launch
- Open weights
- The large model has 303M non-embedding parameters
- Technical report released December 11, 2024
- Released under Apache 2.0
Amazon Nova Canvas
Amazon Nova Canvas is Amazon's image generation model in the first Nova generation, creating images from text and image inputs with editing features such as inpainting, outpainting and background removal.
- Modalities
- text, image → image
- At launch
- API
- Supports inpainting, outpainting and background removal
- Generated images include digital watermarking
- Virtual try-on and visual style options added July 1, 2025
Amazon Nova Lite
Amazon Nova Lite is a low-cost multimodal model in Amazon's first Nova generation that processes text, image and video input, with a context of up to 300K tokens and up to 30 minutes of video per request.
- Modalities
- text, image, video → text
- Context
- 300K context
- At launch
- API
- Context up to 300K tokens
- Processes up to 30 minutes of video in a single request
- Launched in Amazon Bedrock
Amazon Nova Micro
Amazon Nova Micro is the text-only model in Amazon's first Nova generation, focused on very fast, low-cost responses for tasks such as summarization, translation and classification, with a 128K token context window.
- Modalities
- text → text
- Context
- 125K context
- At launch
- API
- Text-only model with a 128K token context window
- Launched in Amazon Bedrock with the rest of the Nova family
- Available in US East (N. Virginia), with more US Regions via cross-Region inference
Amazon Nova Pro
Amazon Nova Pro is the multimodal model balancing accuracy, speed and cost in Amazon's first Nova generation, accepting text, image and video input with up to 300K input tokens.
- Modalities
- text, image, video → text
- Context
- 300K context
- At launch
- API
- Accepts text, image and video input
- Up to 300K input tokens
- Launched in Amazon Bedrock
Amazon Nova Reel
Amazon Nova Reel is Amazon's video generation model in the first Nova generation, producing short videos from text prompts and images with camera control actions such as zooming and camera movement.
- Modalities
- text, image → video
- At launch
- API
- Supports camera control actions such as zooming or moving the camera
- Generated videos include digital watermarking
- Succeeded by Nova Reel 1.1 in April 2025
HunyuanVideo
HunyuanVideo is Tencent's open text-to-video foundation model with over 13 billion parameters, using a multimodal LLM text encoder and joint image-video training, released with inference code and weights.
- Modalities
- text → video
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Over 13 billion parameters
- Recommended output of 720x1280 at 129 frames
- FP8 weights added December 18, 2024
Rerank 3.5
Reranking model with stronger reasoning over complex, constraint-heavy queries and broad data compatibility across long documents, semi-structured data and code.
- Modalities
- text → score
- At launch
- API, Cloud platforms
- Covers 100+ languages with strength in 10 business languages
- Later made available on Azure AI Foundry
- API model ID rerank-v3.5
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 December 2024." The AI Engineering, updated Dec 26, 2024. https://theaiengineering.com/releases/2024/12
HTML link
<a href="https://theaiengineering.com/releases/2024/12">AI models released in December 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