AI models released in February 2025
27 releases from 18 labs in February 2025, including Command R7B Arabic, GPT-4.5 (research preview) and Ideogram 2a. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 27
- Labs
- 18
- Open weights
- 15
February 2025
27 releasesCommand R7B Arabic
Version of Command R7B optimized for Arabic language and culture for enterprises in the Middle East and North Africa, released with open weights.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 7B
- License
- CC-BY-NC-4.0
- At launch
- API, Open weights
- Runs on low-end GPUs, a MacBook or CPUs
- 128K-token context length
- Improves on all Arabic language dimensions versus Command R7B
GPT-4.5 (research preview)
GPT-4.5 is a research preview of OpenAI's largest chat model at the time, built by scaling unsupervised pre-training and post-training. OpenAI cites broader knowledge, better grasp of user intent, and fewer hallucinations, without step-by-step reasoning.
- Modalities
- text, image → text
- Context
- 125K context
- At launch
- Consumer app, API
- API model ID gpt-4.5-preview in the Chat Completions, Assistants, and Batch APIs
- Available first to ChatGPT Pro users, with Plus and Team the following week
- Supports function calling, Structured Outputs, streaming, and image input
- 128,000-token context window per the API model page
Ideogram 2a
Ideogram's fastest and most affordable text-to-image model at launch, optimized for graphic design and photography and released on the website, API and partner platforms.
- Modalities
- text → image
- At launch
- Consumer app, API
- Generates images in 10 seconds, or 5 seconds with 2a Turbo
- Launched with image prices cut by 50% on the website and API
- Offered through partners including Replicate, fal and Freepik
Granite 3.2
Update of the Granite language models that adds optional thinking capabilities to the 2B and 8B instruct models, built on Granite 3.1 Instruct.
- Modalities
- text → text
- Parameters
- 2B and 8B
- License
- Apache-2.0
- At launch
- Open weights
- Reasoning can be toggled on or off
- Built on Granite 3.1 8B Instruct
- Released alongside Granite Vision 3.2 and Granite Guardian 3.2
Granite Guardian 3.2
Smaller Granite Guardian safety models, including a 5B model pruned and healed from Granite Guardian 3.1 8B and a 3B mixture-of-experts model with 800M active parameters.
- Modalities
- text → text
- Parameters
- 5B and 3B-A800M
- License
- Apache-2.0
- At launch
- Open weights
- About 30% of the 8B model's parameters removed for the 5B version
- Detects risks in prompts and responses
- Released under Apache 2.0
Granite Vision 3.2
Compact IBM vision-language model designed for visual document understanding, automating content extraction from tables, charts, infographics, plots and diagrams for enterprise use.
- Modalities
- text, image → text
- Parameters
- 2B
- License
- Apache-2.0
- At launch
- Open weights
- Follows the Granite Vision 3.1 preview from January 2025
- Focused on enterprise document tasks
- Released under Apache 2.0
Phi-4-mini
Phi-4-mini is a 3.8B dense, decoder-only Microsoft language model with grouped-query attention and a 200,000-token vocabulary, designed for speed and efficiency, with support for function calling and long inputs.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 3.8B
- License
- MIT
- At launch
- Open weights, API, Cloud platforms
- Supports sequences up to 128,000 tokens
- 200,000-token vocabulary with grouped-query attention
- Available in Azure AI Foundry, Hugging Face and the NVIDIA API Catalog
Phi-4-multimodal
Phi-4-multimodal is Microsoft's first multimodal Phi language model, a 5.6B model that processes speech, vision and text in a single unified model rather than a pipeline of separate models.
- Modalities
- text, image, audio → text
- Context
- 125K context
- Parameters
- 5.6B
- License
- MIT
- At launch
- Open weights, API, Cloud platforms
- 5.6B parameters with a 128K token context length
- Handles speech, vision and text in one model
- Available in Azure AI Foundry, Hugging Face and the NVIDIA API Catalog
Scribe v1
ElevenLabs' first speech-to-text model, transcribing speech in 99 languages with word-level timestamps, speaker diarization and audio-event tagging in a structured response.
- Modalities
- audio → text
- At launch
- API, Consumer app
- 99 languages
- Word-level timestamps and speaker diarization
- Audio-event tagging in a structured response
olmOCR
Open PDF text extraction toolkit built on a fine-tuned Qwen2-VL-7B-Instruct model, released with weights, fine-tuning data and an efficient inference pipeline.
- Modalities
- text, image → text
- Parameters
- 7B
- License
- Apache-2.0
- At launch
- Open weights
- Built on Qwen2-VL-7B-Instruct
- Released with training and inference code
- Followed by olmOCR 2 in October 2025
QwQ-Max-Preview
QwQ-Max-Preview is a preview of Qwen's QwQ-Max reasoning model, built on Qwen2.5-Max and strong in math, coding, and agent tasks. Qwen said it planned to open-source QwQ-Max and Qwen2.5-Max under Apache 2.0.
- Modalities
- text → text
- At launch
- Consumer app
- Available on Qwen Chat
- Announced alongside plans for a Qwen Chat app
- Qwen also planned smaller open reasoning models such as QwQ-32B
Wan2.1
Wan2.1 is Alibaba's open video generation model family, released with inference code and weights for text-to-video and image-to-video at 1.3B and 14B sizes.
- Modalities
- text, image → video
- Parameters
- 1.3B, 14B
- License
- Apache-2.0
- At launch
- Open weights
- Text-to-video in 1.3B and 14B; image-to-video 14B at 480P and 720P
- First-last-frame-to-video model Wan2.1-FLF2V-14B followed April 17, 2025
- Released under the Apache 2.0 license
Claude 3.7 Sonnet
Claude 3.7 Sonnet is Anthropic's first hybrid reasoning model, able to give near-instant responses or visible step-by-step extended thinking. Anthropic introduced Claude Code, a command line tool for agentic coding, as a research preview alongside it.
- Modalities
- text, image → text
- At launch
- API, Consumer app, Cloud platforms, Enterprise
- API model ID claude-3-7-sonnet-20250219
- API users can set a thinking budget up to the 128K-token output limit
- Available on all Claude plans; extended thinking excluded from the free tier
- Available on Amazon Bedrock and Google Cloud Vertex AI
Moonlight
Moonlight is a 3B-active, 16B-total mixture-of-experts model trained on 5.7T tokens with the Muon optimizer. Moonshot released it to show that Muon scales to large-model training without hyperparameter tuning.
- Modalities
- text → text
- Parameters
- MoE: 16B total, 3B active
- License
- MIT
- At launch
- Open weights
- Moonshot reports Muon is about 2 times more sample efficient than Adam
- Released pretrained, instruction-tuned, and intermediate checkpoints
- Muon implementation open-sourced alongside the model
SmolVLM2
Video-capable SmolVLM models in three sizes that bring visual and video understanding to small devices, with the 500M and 256M versions billed as the smallest video language models.
- Modalities
- text, image, video → text
- Parameters
- 256M, 500M and 2.2B
- License
- Apache-2.0
- At launch
- Open weights
- MLX-ready with Python and Swift APIs from launch
- 2.2B model aimed at vision and video tasks
- Released under Apache 2.0
Grok 3 and Grok 3 mini
Grok 3 is xAI's flagship model trained on the Colossus supercluster, launched in beta alongside Grok 3 mini, with Think reasoning modes refined through large-scale reinforcement learning.
- Modalities
- text → text
- At launch
- Consumer app
- Rolled out to X Premium and Premium+ users on X and Grok.com
- Announcement states a 1M-token context window for Grok 3
- Launched alongside DeepSearch, xAI's first agent
- Grok 3 models became generally available on the API on April 3, 2025
PaliGemma 2 mix
PaliGemma 2 mix is a set of PaliGemma 2 checkpoints fine-tuned on a mixture of tasks, so one model handles captioning, OCR, visual question answering, object detection and segmentation, selected through the prompt.
- Modalities
- text, image → text
- Parameters
- 3B, 10B, 28B
- License
- Gemma Terms of Use
- At launch
- Open weights, Cloud platforms
- Available at 224px and 448px resolutions
- Drop-in upgrade for original PaliGemma mix users
- Weights on Kaggle and Hugging Face; deployable in Vertex Model Garden
R1 1776
A version of DeepSeek-R1 that Perplexity post-trained on about 40,000 multilingual prompts covering topics censored by the Chinese Communist Party, aiming for factual answers while keeping R1's reasoning ability.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Open weights
- Weights released on Hugging Face under the MIT license
- Also offered through Perplexity's Sonar API at launch
- Post-trained with an adapted version of NVIDIA's NeMo 2.0 framework
- Removed from the Perplexity API on August 1, 2025, per the API changelog
Mistral Saba
First of Mistral's specialized regional language models, a 24B model trained on curated data from the Middle East and South Asia for Arabic and Indian-origin languages.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 24B
- At launch
- API, Enterprise
- Particularly strong in South Indian-origin languages such as Tamil
- Available through the API and for local deployment
- API model ID mistral-saba-2502
Step-Audio
Step-Audio is StepFun's open speech interaction family, led by the 130B-parameter Step-Audio-Chat model that unifies speech understanding and generation, released with a 3B text-to-speech model and an audio tokenizer.
- Modalities
- audio, text → text, audio
- Parameters
- 130B
- License
- Apache-2.0
- At launch
- Open weights
- Step-Audio-Chat is a 130B-parameter unified speech model
- Step-Audio-TTS-3B and Step-Audio-Tokenizer released alongside
- Released with the StepEval-Audio-360 multi-turn benchmark
Step-Video-T2V
Step-Video-T2V is StepFun's 30B-parameter open text-to-video model, capable of generating videos of up to 204 frames, released alongside a faster Turbo variant and a technical report.
- Modalities
- text → video
- Parameters
- 30B
- License
- MIT
- At launch
- Open weights
- Generates videos up to 204 frames
- Step-Video-T2V-Turbo released the same day
- Weights on Hugging Face under MIT
MiniMax Image-01
Image-01 is MiniMax's first text-to-image generation model, offered through the MiniMax API platform with an emphasis on prompt adherence, lighting, detailed environments, and realistic people and objects.
- Modalities
- text → image
- At launch
- API
- API release notes list it on February 15, 2025; the announcement post is dated February 28, 2025
- Supports multiple standard aspect ratios
- Later joined by image-01-live for generation from a reference image
DeepHermes-3-Llama-3-8B-Preview
Preview Hermes model from Nous Research that unifies deep reasoning and standard chat modes in a single model, built on Meta's Llama 3 8B.
- Modalities
- text → text
- Parameters
- 8B
- License
- Llama 3 Community License
- At launch
- Open weights
- Switches between deep reasoning and chat modes
- 3B and Mistral 24B preview versions also released
- Built on Llama 3 8B
MiniMax T2V-01-Director and I2V-01-Director
T2V-01-Director and I2V-01-Director are Hailuo text-to-video and image-to-video models that follow camera movement instructions more precisely, reduce movement randomness, and offer preset camera settings for cinematic shots.
- Modalities
- text, image → video
- At launch
- API, Consumer app
- API release notes list the release on February 11, 2025; the announcement post is dated March 3, 2025
- Available on Hailuo AI web and mobile and the MiniMax API
Sonar (built on Llama 3.3 70B)
An updated version of Perplexity's in-house Sonar model, built on Llama 3.3 70B and further trained for answer factuality and readability in Perplexity's default search mode, served on Cerebras inference infrastructure.
- Modalities
- text → text
- Parameters
- 70B (Llama 3.3 70B base)
- At launch
- API, Consumer app
- Available to Perplexity Pro users, who can set it as their default model
- Perplexity reports about 1,200 tokens per second on Cerebras infrastructure
- Also offered via the Sonar API, which did not yet run on Cerebras at launch
Gemini 2.0 Flash-Lite
Gemini 2.0 Flash-Lite was introduced as Google's most cost-efficient model at the time, designed for speed and scale, with a 1 million token context window and multimodal input.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Cloud platforms
- Public preview in Google AI Studio and Vertex AI at launch
- Generally available in the Gemini API and Vertex AI on Feb 25, 2025
- Shut down in the Gemini API on June 1, 2026
Gemini 2.0 Pro
Gemini 2.0 Pro was released as an experimental model that Google described as its best yet for coding performance and complex prompts, with its largest context window at the time of 2 million tokens.
- Modalities
- text, image → text
- Context
- 2M context
- At launch
- API, Cloud platforms, Consumer app
- Experimental model ID gemini-2.0-pro-exp-02-05
- Available in Google AI Studio, Vertex AI and to Gemini Advanced users in the Gemini app
- Able to call tools such as Google Search and code execution
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 February 2025." The AI Engineering, updated Feb 27, 2025. https://theaiengineering.com/releases/2025/02
HTML link
<a href="https://theaiengineering.com/releases/2025/02">AI models released in February 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