AI models released in March 2025
38 releases from 21 labs in March 2025, including Amazon Nova Act, Gen-4 and QVQ-Max. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 38
- Labs
- 21
- Open weights
- 23
March 2025
38 releasesAmazon Nova Act
Amazon Nova Act is an Amazon AGI model trained to perform actions within a web browser, released as a research preview SDK for building agents that complete multi-step browser tasks.
- Modalities
- text → text
- At launch
- API
- Released as a research preview SDK through nova.amazon.com
- Agents can combine natural language commands with Python code and Playwright browser control
- Became an AWS service powered by a custom Nova 2 Lite model in December 2025
Gen-4
Video generation model built for world consistency, generating consistent characters, locations and objects across scenes from visual references and instructions without fine-tuning.
- Modalities
- text, image → video
- At launch
- Consumer app
- Available to paid plans from April 1, 2025
- Regenerates elements from multiple perspectives within scenes
- Released with short films made to test narrative capabilities
QVQ-Max
QVQ-Max is the first official version of Qwen's visual reasoning model, following the exploratory QVQ-72B-Preview. It understands images and videos and reasons over them to solve problems.
- Modalities
- text, image, video → text
- At launch
- Consumer app
- Accuracy improves as the maximum thinking length increases, per Qwen
- Covers math problems, everyday questions, code, and creative tasks
- Available to try on Qwen Chat
Qwen2.5-Omni
Qwen2.5-Omni is Qwen's end-to-end multimodal model that perceives text, images, audio, and video and responds with text and natural speech in a streaming manner. It uses a Thinker-Talker architecture.
- Modalities
- text, image, audio, video → text, audio
- Parameters
- 7B
- License
- Apache-2.0
- At launch
- Consumer app, Open weights
- Released as Qwen2.5-Omni-7B; a 3B version followed on April 30, 2025
- Available on Qwen Chat
- Weights under the Apache 2.0 license
Ideogram 3.0
Ideogram image model focused on photorealism, legible text and creative designs, and the only Ideogram model offering style presets, style reference images, negative prompts and color palettes.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Style references guide the look of generations
- Transparent-background variant available
- Remains available in the app and API alongside newer models
xLAM-2
Second generation of Salesforce's large action models for function calling and AI agents, adding multi-turn conversation and tool use trained with the APIGen-MT data pipeline.
- Modalities
- text → text
- Parameters
- 1B, 3B, 8B, 32B and 70B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Released in 1B, 3B, 8B, 32B and 70B sizes on Llama and Qwen backbones
- Llama-based models supported 128k context; Qwen-based models 32k, extendable to 128k
- Released for research use under CC BY-NC 4.0
DeepSeek-V3-0324
DeepSeek-V3-0324 is an updated DeepSeek-V3 checkpoint with a major boost in reasoning, stronger front-end development skills, and smarter tool use. It also moved the V3 weights to the MIT License.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Replaced V3 behind the deepseek-chat API model on March 24, 2025
- DeepSeek recommends it for non-complex reasoning tasks with DeepThink turned off
- Weights on Hugging Face under the MIT License
Gemini 2.5 Pro
Gemini 2.5 Pro was the first Gemini 2.5 model, a thinking model that reasons through its thoughts before responding, launched as an experimental release with a 1 million token context window.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms, Consumer app
- Available at launch in Google AI Studio and to Gemini Advanced users in the Gemini app
- Paid preview gemini-2.5-pro-preview-03-25 followed on Apr 4, 2025
- Stable gemini-2.5-pro became generally available on June 17, 2025
- Accepts audio, images, video, text and PDF input
GPT-4o image generation (gpt-image-1)
4o image generation builds image generation natively into GPT-4o, with improved text rendering, multi-turn generation, and instruction following. It became ChatGPT's default image generator and later reached the API as gpt-image-1.
- Modalities
- text, image → image
- At launch
- Consumer app
- Rolled out to Free, Plus, Pro, and Team users in ChatGPT and also in Sora
- Released in the API as gpt-image-1 on April 23, 2025
- Replaced DALL-E as ChatGPT's default image generator; DALL-E remained via a GPT
TxGemma
TxGemma is a collection of open models fine-tuned from Gemma 2 for therapeutics development, covering classification, regression and generation tasks across drug discovery. Predict variants target narrow tasks, while chat variants explain their reasoning.
- Modalities
- text → text
- Parameters
- 2B, 9B, 27B
- License
- Health AI Developer Foundations Terms of Use
- At launch
- Open weights, Cloud platforms
- Fine-tuned on 7 million training examples
- Predict models at 2B, 9B and 27B; chat models at 9B and 27B
- Available on Vertex AI Model Garden and Hugging Face
Qwen2.5-VL-32B
Qwen2.5-VL-32B-Instruct is a 32B vision-language model in the Qwen2.5-VL series, further optimized with reinforcement learning. Its responses align more closely with human preferences.
- Modalities
- text, image, video → text
- Context
- 32K context
- Parameters
- 32B
- License
- Apache-2.0
- At launch
- Open weights
- Stronger mathematical reasoning and fine-grained image understanding
- Released under the Apache 2.0 license
- Weights on Hugging Face and ModelScope
gpt-4o-mini-tts
gpt-4o-mini-tts is a text-to-speech model built on GPT-4o mini that developers can instruct on how to speak, such as tone and style, not only on what to say. It is limited to preset synthetic voices.
- Modalities
- text → audio
- At launch
- API
- API model ID gpt-4o-mini-tts in the text-to-speech API
- Steerable delivery, for example a sympathetic customer service tone
- Released with an Agents SDK integration for building voice agents
gpt-4o-transcribe and gpt-4o-mini-transcribe
gpt-4o-transcribe and gpt-4o-mini-transcribe are speech-to-text models built on GPT-4o and GPT-4o mini. OpenAI says they transcribe more accurately than its Whisper models, especially with accents, noise, and varying speech speeds.
- Modalities
- audio, text → text
- Context
- 16K context
- At launch
- API
- API model IDs gpt-4o-transcribe and gpt-4o-mini-transcribe in the speech-to-text API
- Trained with reinforcement learning and audio-centric midtraining, per OpenAI
- 16,000-token context window per the API model pages
Hunyuan3D-2mini and Hunyuan3D-2mv
Tencent extended Hunyuan3D 2.0 with Hunyuan3D-2mini, a smaller 0.6B image-to-shape model, and Hunyuan3D-2mv, a 1.1B shape model conditioned on multiview images, followed by turbo distilled versions.
- Modalities
- image → 3d
- Parameters
- 0.6B (2mini), 1.1B (2mv)
- License
- Tencent Hunyuan 3D 2.0 Community License Agreement
- At launch
- Open weights
- Turbo step-distilled versions and FlashVDM released March 19, 2025
- Guidance-distilled Fast variants released with the base models
Isaac GR00T N1
NVIDIA's first open, customizable foundation model for generalized humanoid robot reasoning and skills, using a dual-system design for planning and fast motor actions.
- Modalities
- text, image → action
- At launch
- Open weights
- System 2 vision language model plans; System 1 generates continuous actions
- First of a planned family of robot foundation models
- Followed by GR00T N1.5, N1.6 and N1.7
Llama Nemotron Nano and Super
Open reasoning models post-trained by NVIDIA from Meta Llama models, with Nano derived from Llama 3.1 8B and Super from Llama 3.3 70B through neural architecture search.
- Modalities
- text → text
- Parameters
- 8B and 49B
- License
- NVIDIA Open Model License
- At launch
- Open weights, API
- Offered as hosted APIs on build.nvidia.com and as weights on Hugging Face
- A larger Ultra model was announced for multi-GPU servers
- Supports reasoning and chat use cases
Stable Virtual Camera
Multi-view video generation model that turns 2D images into immersive videos with 3D camera control, keeping 3D consistency across sequences up to 1,000 frames.
- Modalities
- image → video
- License
- Stability AI Non-Commercial Research Community License
- At launch
- Open weights
- Research release under a non-commercial license
- Weights on Hugging Face and code on GitHub
- Supports seamless loops and revisiting viewpoints
Mistral Small 3.1
Successor to Mistral Small 3 that adds image understanding and expands the context window to 128K tokens while keeping the 24B Apache 2.0 open-weight design.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Base and instruct checkpoints released
- Improved text performance over Mistral Small 3
- API model ID mistral-small-2503
Step-Video-TI2V
Step-Video-TI2V is a 30B-parameter text-driven image-to-video model built on Step-Video-T2V, generating videos of up to 102 frames from an input image and text prompt.
- Modalities
- text, image → video
- Parameters
- 30B
- License
- MIT
- At launch
- Open weights
- Generates videos up to 102 frames
- Released with the Step-Video-TI2V-Eval benchmark and ComfyUI integration
- Weights on Hugging Face under MIT
ERNIE 4.5
ERNIE 4.5 is Baidu's new-generation native multimodal foundation model, jointly modeling text, images, audio, and video, with stronger language understanding, generation, logic, and memory and reduced hallucination.
- Modalities
- text, image, audio, video → text
- At launch
- Consumer app, API
- Free on the ERNIE Bot website at launch
- API available to enterprises and developers on Baidu AI Cloud Qianfan
- Uses FlashMask dynamic attention masking and heterogeneous multimodal experts
ERNIE X1
ERNIE X1 is Baidu's deep-thinking reasoning model, released with ERNIE 4.5 and described as the first deep-thinking model to use tools autonomously, including search, document Q&A, image understanding, and a code interpreter.
- Modalities
- text, image → text
- At launch
- Consumer app
- Free to users on the ERNIE Bot website at launch
- Qianfan API availability announced as coming soon
- Trained with progressive reinforcement learning and end-to-end chain-of-thought and action training
Command A
Enterprise generative model optimized for tool use, RAG, agents and multilingual work that aims for strong performance on minimal hardware.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- 111B
- License
- CC-BY-NC-4.0
- At launch
- API, Open weights
- 256K-token context length
- Supports 23 languages
- Research weights released as c4ai-command-a-03-2025
OLMo 2 32B
Largest model in the OLMo 2 family and a fully open model trained on up to 6 trillion tokens and post-trained with Tülu 3.1.
- Modalities
- text → text
- Parameters
- 32B
- License
- Apache-2.0
- At launch
- Open weights
- Trained with the new OLMo-core codebase
- Trained on Google Cloud's Augusta AI Hypercomputer
- Released under Apache 2.0
Gemini 2.0 Flash native image generation
Google opened Gemini 2.0 Flash's native image output to developer experimentation, letting the model generate and edit images conversationally alongside text after first offering the capability to trusted testers in December 2024.
- Modalities
- text, image → text, image
- At launch
- API
- Experimental model available in Google AI Studio and the Gemini API in all supported regions
- Supports multi-turn conversational image editing
- Preview model gemini-2.0-flash-preview-image-generation followed on May 7, 2025
Gemma 3
Gemma 3 is Google's third generation of open models in 1B, 4B, 12B and 27B sizes, adding image and short video input to the 4B and larger models. It expands the context window to 128K tokens.
- Modalities
- text, image, video → text
- Context
- 128K context
- Parameters
- 1B, 4B, 12B, 27B
- License
- Gemma Terms of Use
- At launch
- Open weights, API, Cloud platforms
- 128K-token context for 4B, 12B and 27B; 32K for 1B
- Built-in support for over 35 languages and pretrained support for over 140
- Function calling and structured output
- Official quantized versions released
OASIS-code-embedding-1.5B
OASIS-code-embedding-1.5B is a code embedding model from Kuaishou's Kwaipilot team, designed to improve semantic code retrieval and search across programming contexts.
- Modalities
- text → text
- Parameters
- 1.5B
- License
- MIT
- At launch
- Open weights
- Built for code retrieval and code search systems
- Released with a preprint on the same day
- Weights on Hugging Face under MIT
ShieldGemma 2
ShieldGemma 2 is a 4B image safety classifier built on Gemma 3, released alongside it. It labels images against safety policies so developers can filter inputs and outputs of image models.
- Modalities
- text, image → text
- Parameters
- 4B
- License
- Gemma Terms of Use
- At launch
- Open weights
- Covers dangerous content, sexually explicit content and violence
- Trained on Gemma 3's 4B instruction-tuned checkpoint
- Outputs Yes/No probability scores per policy
GPT-4o Search Preview and GPT-4o mini Search Preview
GPT-4o Search Preview and GPT-4o mini Search Preview are fine-tuned web search models offered directly in the Chat Completions API, launched alongside the Responses API and its built-in web search tool.
- Modalities
- text → text
- Context
- 125K context
- At launch
- API
- API model IDs gpt-4o-search-preview and gpt-4o-mini-search-preview
- Web search in the API uses the same model as ChatGPT search, per OpenAI
- Responses include inline citations linking to sources
- 128,000-token context window per the API model pages
Reka Flash 3
An open research preview of Reka's 21B general-purpose reasoning model, trained from scratch and refined with RLOO reinforcement learning, which thinks inside reasoning tags and supports budget forcing to cap thinking length.
- Modalities
- text → text
- Parameters
- 21B
- License
- Apache-2.0
- At launch
- Consumer app, Open weights
- Released with a 32k context length
- Weights on Hugging Face under Apache-2.0, in a Llama-compatible format
- Primarily an English model; available to try in Reka Space
- An improved version powers Reka Nexus, Reka's platform for AI workers
Gemini Embedding (gemini-embedding-001)
Gemini Embedding is Google's first text embedding model trained on the Gemini model itself, released experimentally in March 2025 and as the stable gemini-embedding-001 in July 2025 for retrieval, classification and RAG.
- Modalities
- text → embedding
- Context
- 2K context
- At launch
- API, Cloud platforms
- Supports over 100 languages
- Default 3072-dimension output that can be scaled down with Matryoshka Representation Learning
- The experimental gemini-embedding-exp-03-07 accepted 8K input tokens; the GA model supports 2,048
- Generally available in the Gemini API and Vertex AI on July 14, 2025
Chirp 3
Chirp 3 is Google Cloud's latest generation of speech models, covering Chirp 3: HD voices for text-to-speech and Chirp 3: Transcription for speech recognition. HD voices expanded to 8 speakers in 31 locales in March 2025.
- Modalities
- text, audio → audio, text
- At launch
- API, Cloud platforms
- Chirp 3: HD voices with 8 speakers and 31 locales generally available from April 2, 2025
- Instant Custom Voice creates a custom voice from 10 seconds of audio (allowlist)
- Chirp 3: Transcription generally available from October 13, 2025 with diarization and automatic language detection
HunyuanVideo-I2V
HunyuanVideo-I2V is an open image-to-video generation model based on HunyuanVideo, released with inference code, weights, and LoRA training code for customizable video effects.
- Modalities
- text, image → video
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Parallel inference via xDiT added March 13, 2025
- LoRA effects examples include hair growth and embrace
Jamba 1.6
Jamba Large 1.6 and Jamba Mini 1.6, an update of AI21's open hybrid SSM-Transformer family aimed at long-context enterprise deployment.
- Modalities
- text → text
- Context
- 250K context
- License
- Jamba Open Model License
- At launch
- Open weights, API, Enterprise
- 256K-token context window
- Deployable privately in VPC or on-premises
- Weights on Hugging Face
Mistral OCR
Optical character recognition API that reads images and PDFs and returns ordered, interleaved text and images, including tables and equations, for use in RAG pipelines.
- Modalities
- text, image → text
- At launch
- API, Consumer app
- Became the default document understanding model in Le Chat
- API model ID mistral-ocr-2503
- Since deprecated in favor of newer OCR models
QwQ-32B
QwQ-32B is Qwen's 32B reasoning model trained with scaled reinforcement learning, which Qwen says performs comparably to the much larger DeepSeek-R1. It integrates agent capabilities for reasoning with tools.
- Modalities
- text → text
- Context
- 128K context
- Parameters
- 32B
- License
- Apache-2.0
- At launch
- Consumer app, Open weights
- Full 131,072-token context length
- Accessible via Qwen Chat
- Weights on Hugging Face and ModelScope under the Apache 2.0 license
LTX-Video 0.9.5
LTX-Video checkpoint with improved quality, keyframe and video extension support, higher resolutions, better prompt understanding and an improved VAE, moving to a license that permits commercial use.
- Modalities
- text, image → video
- Parameters
- 2B
- License
- OpenRAIL-M
- At launch
- Open weights, Consumer app
- New OpenRAIL-M license allowing commercial use
- Keyframes and video extension
- Online web app in LTX Studio
Aya Vision
Multilingual vision-language models from Cohere For AI that extend multimodal capabilities to 23 languages for captioning, visual question answering and translation.
- Modalities
- text, image → text
- Parameters
- 8B and 32B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Released in 8B and 32B sizes
- Uses synthetic annotations and multimodal model merging
- Weights published under a non-commercial license
CogView4
CogView4 is Z.ai's open-weights text-to-image model with 6B parameters that supports native Chinese input and Chinese text-to-image generation. Z.ai released it in a diffusers-compatible version with an online demo.
- Modalities
- text → image
- Parameters
- 6B
- License
- Apache-2.0
- At launch
- Open weights
- Native Chinese prompt input
- Diffusers version open-sourced on March 4, 2025
- Weights under the Apache 2.0 license
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 2025." The AI Engineering, updated Mar 31, 2025. https://theaiengineering.com/releases/2025/03
HTML link
<a href="https://theaiengineering.com/releases/2025/03">AI models released in March 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