AI models released in January 2025
25 releases from 15 labs in January 2025, including OpenAI o3-mini, Mistral Small 3 and SFR-Embedding-Code. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 25
- Labs
- 15
- Open weights
- 13
January 2025
25 releasesOpenAI o3-mini
o3-mini is a small, cost-efficient reasoning model optimized for STEM tasks such as science, math, and coding. It replaced o1-mini in ChatGPT and offers low, medium, and high reasoning effort in the API.
- Modalities
- text → text
- Context
- 200K context
- At launch
- Consumer app, API
- API model ID o3-mini (snapshot o3-mini-2025-01-31)
- First OpenAI small reasoning model with function calling, Structured Outputs, and developer messages
- Free ChatGPT users could try it via 'Reason'; paid users also got o3-mini-high
- No vision support; 200,000-token context window per the API model page
Mistral Small 3
Latency-optimized 24B model released under Apache 2.0 with pretrained and instruct checkpoints, marking Mistral's return to Apache licensing for general-purpose models.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Uses far fewer layers than comparable models to cut latency
- Not trained with RL or synthetic data
- API model ID mistral-small-2501
SFR-Embedding-Code
Family of embedding models for code and text retrieval in 400M, 2B and 7B sizes, with the 400M and 2B models released openly.
- Modalities
- text → embedding
- Parameters
- 400M and 2B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- The 7B model ranked first on the CoIR code retrieval benchmark at announcement
- The 400M model was designed for faster, lower-cost inference, including on CPUs
- Open models released for research use under CC BY-NC 4.0
Tülu 3 405B
Application of the fully open Tülu 3 post-training recipe to the Llama 3.1 405B base model, used to stress-test reinforcement learning with verifiable rewards at scale.
- Modalities
- text → text
- Parameters
- 405B
- License
- Llama 3.1 Community License
- At launch
- Open weights
- Trained with reinforcement learning from verifiable rewards
- Served with vLLM using 16-way tensor parallelism during RL
- Built on Llama 3.1 405B
Qwen2.5-Max
Qwen2.5-Max is a large-scale mixture-of-experts model pre-trained on more than 20T tokens and post-trained with supervised fine-tuning and RLHF. Qwen released its API through Alibaba Cloud.
- Modalities
- text → text
- At launch
- API, Consumer app
- Mixture-of-experts architecture
- API available through Alibaba Cloud Model Studio
- Available to try on Qwen Chat
Janus-Pro
Janus-Pro is an advanced version of Janus, DeepSeek's autoregressive framework that unifies multimodal understanding and image generation by decoupling visual encoding into separate pathways. It significantly improves both understanding and visual generation.
- Modalities
- text, image → text, image
- Parameters
- 1B, 7B
- License
- DeepSeek Model License
- At launch
- Open weights
- Built on DeepSeek-LLM-1.5b-base and DeepSeek-LLM-7b-base
- Single unified transformer for image understanding and text-to-image generation
- Code under the MIT License; model use under the DeepSeek Model License
Qwen2.5-1M
Qwen2.5-1M is the first set of open Qwen models upgraded to handle 1M-token contexts, released as Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M with a vLLM-based inference framework.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- 7B, 14B
- License
- Apache-2.0
- At launch
- Open weights
- Context length of 1,010,000 tokens with 8,192-token generation
- Inference framework with sparse attention processes 1M-token inputs 3 to 7 times faster
- Weights under the Apache 2.0 license
Qwen2.5-VL
Qwen2.5-VL is Qwen's vision-language model series released in 3B, 7B, and 72B sizes as base and instruct models. The flagship Qwen2.5-VL-72B-Instruct covers document understanding, video understanding, and visual agent tasks.
- Modalities
- text, image, video → text
- Context
- 32K context
- Parameters
- 3B, 7B, 72B
- License
- Apache-2.0 (7B); Qwen License Agreement (72B); Qwen Research License Agreement (3B)
- At launch
- Consumer app, Open weights
- Qwen2.5-VL-72B-Instruct available on Qwen Chat
- Qwen2.5-VL-3B targets edge AI
- Weights on Hugging Face and ModelScope
Hunyuan-7B-0124
Hunyuan-7B-0124 is an open 7B dense language model released in pretrained and instruct versions, using improved data allocation and training, grouped query attention, and a context window extended to 256K.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- 7B
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Released as Hunyuan-7B-Pretrain-0124 and Hunyuan-7B-Instruct-0124
- Hugging Face format with a vLLM inference backend
Computer-Using Agent (computer-use-preview)
Computer-Using Agent (CUA) combines GPT-4o's vision with reasoning trained through reinforcement learning to operate graphical interfaces from screenshots using a virtual mouse and keyboard. It launched powering the Operator research preview.
- Modalities
- text, image → text
- Context
- 8K context
- At launch
- Consumer app
- Launched in Operator for ChatGPT Pro users in the U.S.
- Released in the API as computer-use-preview with the Responses API on March 11, 2025
- Asks for user confirmation on sensitive actions such as logins and CAPTCHAs
- API model page lists an 8,192-token context window
SmolVLM 256M and 500M
Smaller additions to the SmolVLM family, with the 256M model presented as the smallest vision-language model in the world at release.
- Modalities
- text, image → text
- Parameters
- 256M and 500M
- License
- Apache-2.0
- At launch
- Open weights
- Released with ColSmolVLM retrieval models
- Base and instruct versions
- Released under Apache 2.0
Doubao-1.5-pro
Doubao-1.5-pro is ByteDance's sparse mixture-of-experts flagship model, built with an integrated training and inference design that uses few active parameters, with upgraded vision and end-to-end speech capabilities.
- Modalities
- text, image → text
- Context
- 250K context
- At launch
- API
- Offered on Volcano Engine in 32K and 256K context versions
- Released alongside Doubao-1.5-lite and Doubao-1.5-vision-pro
- Uses a 2.4B-parameter Doubao ViT with native dynamic resolution
- Post-training used no data from other models
Hunyuan3D 2.0
Hunyuan3D 2.0 is an open large-scale 3D asset generation system that pairs a shape generation model with a texture synthesis model to produce high-resolution textured 3D assets from images.
- Modalities
- image → 3d
- License
- Tencent Hunyuan 3D 2.0 Community License Agreement
- At launch
- Open weights, Consumer app
- Hunyuan3D-DiT-v2-0 shape model has 1.1B parameters and Hunyuan3D-Paint-v2-0 has 1.3B
- Launched alongside the Hunyuan3D Studio website
Sonar and Sonar Pro (API)
Perplexity's Sonar API models with built-in real-time web search and citations: Sonar for fast, lightweight answers and Sonar Pro for multi-step queries with more citations and a larger context window.
- Modalities
- text → text
- At launch
- API
- API model IDs sonar and sonar-pro; Sonar Pro generally available to all developers at launch
- Model docs listed 127k context for sonar and 200k for sonar-pro, with an 8k output cap on sonar-pro
- Replaced the llama-3.1-sonar online models, which were retired on February 22, 2025
- JSON mode and search domain filters opened to select usage tiers
DeepSeek-R1
DeepSeek-R1 is an open-weights reasoning model trained with large-scale reinforcement learning in post-training, released with a technical report. DeepSeek also released DeepSeek-R1-Zero, trained with reinforcement learning without supervised fine-tuning.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 671B total, 37B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Code and weights under the MIT License; API outputs may be used for fine-tuning and distillation
- Available in the API as deepseek-reasoner and as DeepThink on chat.deepseek.com
- 128K context length per the model card
DeepSeek-R1-Distill
DeepSeek-R1-Distill is a set of six dense models fine-tuned on reasoning samples generated by DeepSeek-R1, based on Qwen2.5 and Llama 3 checkpoints. They bring R1-style reasoning to sizes from 1.5B to 70B parameters.
- Modalities
- text → text
- Parameters
- 1.5B, 7B, 8B, 14B, 32B, 70B
- License
- MIT
- At launch
- Open weights
- Fine-tuned with 800K samples curated with DeepSeek-R1
- Qwen-based models derive from Qwen2.5 (Apache 2.0); Llama-based models from Llama 3.1 8B and Llama 3.3 70B
- Released the same day as DeepSeek-R1
Doubao Realtime Voice Model
The Doubao Realtime Voice Model is an end-to-end speech understanding and generation model that replaces cascaded pipelines for spoken conversation, offering expressive voice control, emotional responsiveness, low latency, and smooth interruption.
- Modalities
- audio → audio
- At launch
- Consumer app
- Fully available in the Doubao app from version 7.2.0
- Integrates speech and text modalities natively
- Supports natural interruptions during dialogue
Kimi k1.5
Kimi k1.5 is Moonshot AI's multimodal reasoning model trained with reinforcement learning, which Moonshot describes as o1-level. Its training scaled the reinforcement learning context window to 128K and used partial rollouts for efficiency.
- Modalities
- text, image → text
- At launch
- Consumer app
- Jointly trained on text and vision data
- Reports long chain-of-thought and short chain-of-thought results
- Moonshot said the service would come to kimi.ai soon after the January 20, 2025 report
LFM-7B
A 7.7B-parameter Liquid Foundation Model tuned for chat, instruction following, and multilingual use, with Liquid AI highlighting English, Arabic, and Japanese, aimed at private enterprise, on-premises, and on-device deployment.
- Modalities
- text → text
- Parameters
- 7.7B
- At launch
- API, Consumer app, Cloud platforms, Enterprise
- Supports English, Spanish, French, German, Chinese, Arabic, Japanese, and Korean
- Effective context length of 32k tokens validated with RULER
- Offered on Liquid Playground, OpenRouter, and AWS Marketplace
- Weights not released; on-premises customization stack sold to enterprises
MiniMax-Text-01 and MiniMax-VL-01
MiniMax-01 is an open-weight series pairing the MiniMax-Text-01 language model with the MiniMax-VL-01 vision-language model, built on a hybrid of Lightning Attention, softmax attention, and mixture-of-experts.
- Modalities
- text, image → text
- Context
- 4M context
- Parameters
- MoE: 456B total, 45.9B active
- License
- MiniMax Model License
- At launch
- API, Consumer app, Open weights
- 456B total parameters with 45.9B activated per token
- Trained at a 1M-token context and handles up to 4M tokens at inference
- VL-01 adds a 303M-parameter vision transformer to the Text-01 base
- Model weights under the MiniMax Model License; code under MIT
Ray2
Second-generation Luma video model built on a new multimodal architecture trained with ten times the compute of Ray1, available in Dream Machine and later with learnable camera motion concepts.
- Modalities
- text, image → video
- At launch
- Consumer app
- Ten times the compute of Ray1
- Camera Motion Concepts added March 31, 2025
- Announced for Adobe Firefly in April 2025
Codestral 25.01
Update of Codestral with a more efficient architecture and improved tokenizer that generates and completes code about twice as fast as the original, aimed at fill-in-the-middle use.
- Modalities
- text → text
- At launch
- API
- Supports fill-in-the-middle, code correction and test generation
- Optimized for low-latency, high-frequency coding tasks
- API model ID codestral-2501
MiniMax I2V-01-live
I2V-01-live is a Hailuo image-to-video model in the I2V family designed to animate 2D illustrations, with smoother, more vivid motion, greater stability, and subtle expressions across a range of art styles.
- Modalities
- text, image → video
- At launch
- API, Consumer app
- Aimed at bringing 2D illustrations and characters to life
- Available in Hailuo AI and through the MiniMax API
MiniMax S2V-01
S2V-01 is Hailuo's subject reference video model, generating videos that keep a person's facial features and identity consistent across camera angles and movement from a single reference image.
- Modalities
- text, image → video
- At launch
- API, Consumer app
- Needs only one reference image
- Posture, expression, and lighting can be adjusted with text prompts
Cosmos World Foundation Models
Platform of generative world foundation models for physical AI, with diffusion and autoregressive models, tokenizers and guardrails for robotics and autonomous vehicles.
- Modalities
- text, image, video → video
- License
- NVIDIA Open Model License
- At launch
- Open weights, API
- Announced at CES 2025
- Weights on Hugging Face and NVIDIA NGC under NVIDIA's open model license
- Previewable in the NVIDIA API catalog
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 January 2025." The AI Engineering, updated Jan 31, 2025. https://theaiengineering.com/releases/2025/01
HTML link
<a href="https://theaiengineering.com/releases/2025/01">AI models released in January 2025</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering