AI models released in April 2025
40 releases from 21 labs in April 2025, including Amazon Nova Premier, DeepSeek-Prover-V2 and Gen-4 References (Gen-4 Image). Every entry is dated to the lab's official announcement and links to it.
- Releases
- 40
- Labs
- 21
- Open weights
- 19
April 2025
40 releasesAmazon Nova Premier
Amazon Nova Premier is the most capable model in the first Amazon Nova generation, built for complex tasks and to serve as a teacher for distilling smaller custom models, with a one-million-token context window.
- Modalities
- text, image, video → text
- Context
- 1M context
- At launch
- API
- Accepts text, image and video input; audio is not supported
- One-million-token context window
- First announced as coming on December 3, 2024
DeepSeek-Prover-V2
DeepSeek-Prover-V2 is an open-weights model for formal theorem proving in Lean 4, trained with cold-start data from a recursive proof-search pipeline powered by DeepSeek-V3. It comes in 7B and 671B sizes.
- Modalities
- text → text
- Parameters
- 7B, 671B
- License
- DeepSeek Model License
- At launch
- Open weights
- DeepSeek-Prover-V2-671B is trained on top of DeepSeek-V3-Base
- DeepSeek-Prover-V2-7B builds on Prover-V1.5-Base with context extended to 32K tokens
- Use is subject to the DeepSeek model license
Gen-4 References (Gen-4 Image)
Gen-4 image generation model that uses reference images to create consistent characters, locations and other elements, launched as Gen-4 References for paid plans.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Extended to Free plan users on May 7, 2025
- Gen-4 Image API launched May 16, 2025
- Improved object consistency update on June 25, 2025
MiMo-7B
MiMo-7B is Xiaomi's first open-source language model series, built for math and code reasoning and released as base, SFT, RL-Zero, and RL checkpoints with pre-training and post-training tailored to reasoning.
- Modalities
- text → text
- Parameters
- 7B
- License
- MIT
- At launch
- Open weights
- Base model pre-trained on about 25 trillion tokens with multiple-token prediction
- MiMo-7B-RL-0530 update on May 30, 2025 scaled SFT data and extended the RL window to 48K
- Weights under the MIT license
Phi-4-mini-reasoning
Phi-4-mini-reasoning is a 3.8B open Microsoft model for math reasoning, fine-tuned from Phi-4-mini with synthetic data generated by DeepSeek-R1, and aimed at edge, mobile and educational applications.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 3.8B
- License
- MIT
- At launch
- Open weights, API
- 3.8B parameters with a 128K token context length
- Fine-tuned on synthetic data generated by DeepSeek-R1
- ONNX version for Copilot+ PCs made available May 15, 2025
Phi-4-reasoning and Phi-4-reasoning-plus
Phi-4-reasoning is a 14B open-weight reasoning model fine-tuned from Phi-4 on reasoning demonstrations, released with Phi-4-reasoning-plus, which adds reinforcement learning to use more inference-time compute.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 14B
- License
- MIT
- At launch
- Open weights, API
- Both models have 14B parameters and a 32K token context length
- Phi-4-reasoning-plus is further trained with reinforcement learning
- Available on Azure AI Foundry and Hugging Face under the MIT license
Llama Guard 4
Llama Guard 4 is a 12B natively multimodal safety classifier that moderates text and multi-image inputs and model responses, acting as a unified safeguard across modalities.
- Modalities
- text, image → text
- Parameters
- 12B
- License
- Llama 4 Community License
- At launch
- Open weights, API
- Dense model pruned from the Llama 4 Scout pretrained model
- Classifies content against 14 MLCommons-based hazard categories
- Also offered on the Llama API, then in limited preview
Llama Prompt Guard 2
Llama Prompt Guard 2 updates Meta's prompt attack classifier with an improved 86M model and a smaller 22M model for detecting prompt injection and jailbreak attempts.
- Modalities
- text → text
- Context
- 1K context
- Parameters
- 86M, 22M
- License
- Llama 4 Community License
- At launch
- Open weights
- The 22M model cuts latency and compute costs by up to 75% versus the 86M model
- 86M is based on mDeBERTa-base; 22M on DeBERTa-xsmall
- Both support a 512-token context window
Qwen3
Qwen3 is Alibaba's open-weights model family of two mixture-of-experts models and six dense models from 0.6B to 32B parameters. The models offer hybrid thinking and non-thinking modes with user control over how much reasoning a task receives.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- MoE: 235B total, 22B active (flagship); 0.6B to 32B dense
- License
- Apache-2.0
- At launch
- API, Consumer app, Open weights
- Pre-trained on about 36T tokens covering 119 languages and dialects
- Flagship Qwen3-235B-A22B supports 32,768 tokens natively and 131,072 with YaRN
- Weights under the Apache 2.0 license
ERNIE 4.5 Turbo
ERNIE 4.5 Turbo is a faster multimodal model derived from ERNIE 4.5, launched at Baidu Create 2025 with improvements in hallucination reduction, logical reasoning, and coding.
- Modalities
- text, image, video → text
- At launch
- API
- Launched at Baidu Create 2025 in Wuhan
- Enhanced multimodal capabilities, launched alongside ERNIE X1 Turbo
- Serves as the base for the ERNIE X1 Turbo reasoning model
ERNIE X1 Turbo
ERNIE X1 Turbo is a deep-thinking model built on ERNIE 4.5 Turbo, launched at Baidu Create 2025 with a more advanced chain of thought and stronger question answering, writing, reasoning, tool use, and multimodal abilities.
- Modalities
- text, image → text
- At launch
- API
- Launched with ERNIE 4.5 Turbo at Create 2025 in Wuhan
- Combines thinking with tool calls in a compound chain of thought and action
- Built on ERNIE 4.5 Turbo, which is derived from ERNIE 4.5
Kimi-Audio
Kimi-Audio is an open-source audio foundation model for audio understanding, generation, and conversation in one framework. It handles tasks such as speech recognition, audio question answering, captioning, emotion recognition, and end-to-end speech conversation.
- Modalities
- text, audio → text, audio
- Parameters
- 7B
- License
- MIT
- At launch
- Open weights
- Pre-trained on more than 13 million hours of audio and text data
- Hybrid input of continuous acoustic and discrete semantic tokens
- Kimi-Audio-7B-Instruct released April 25, 2025; base Kimi-Audio-7B weights followed April 27
Step1X-Edit
Step1X-Edit is StepFun's open-source general image editing model that follows natural-language editing instructions, released with inference code, weights, a technical report, and the GEdit-Bench evaluation benchmark.
- Modalities
- text, image → image
- License
- Apache-2.0
- At launch
- Open weights
- Released with the GEdit-Bench image editing benchmark
- Updated v1.1 weights followed on July 9, 2025
- Weights on Hugging Face and ModelScope under Apache-2.0
Lyria 2
Lyria 2 is Google DeepMind's music generation model for high-fidelity, professional-grade audio, introduced with an expanded Music AI Sandbox. It later became generally available to enterprises on Vertex AI and to creators through YouTube Shorts.
- Modalities
- text → audio
- At launch
- Consumer app, Cloud platforms
- All Lyria 2 output is watermarked with SynthID
- Music AI Sandbox access expanded to more U.S. musicians, producers and songwriters
- Generally available on Vertex AI from May 20, 2025
Lyria RealTime
Lyria RealTime is Google DeepMind's interactive music generation model that lets users create, perform and steer a continuous music stream in real time, mixing genres and blending styles as it plays. It powers MusicFX DJ.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Output is watermarked with SynthID
- Released in the Gemini API and AI Studio as lyria-realtime-exp on May 20, 2025
- Its open-weights counterpart is Magenta RealTime
GLM-ASR
GLM-ASR is Z.ai's speech recognition model on the BigModel API, which uses context to transcribe audio into idiomatic, readable text. Z.ai says it performs better in noisy environments than its earlier models.
- Modalities
- audio → text
- At launch
- API
- Supports Chinese, English, and several regional Chinese dialects
- Not disrupted by non-speech noise
- Available through the BigModel API platform
Gemma 3 QAT
Gemma 3 QAT is a set of quantization-aware trained int4 Gemma 3 checkpoints that cut memory needs sharply while preserving quality. The 27B model's weights shrink from 54GB in BF16 to 14.1GB, fitting a single consumer GPU.
- Modalities
- text, image → text
- Context
- 128K context
- Parameters
- 1B, 4B, 12B, 27B
- License
- Gemma Terms of Use
- At launch
- Open weights
- Covers the 27B, 12B, 4B and 1B Gemma 3 sizes
- 27B fits on a single 24GB card such as the NVIDIA RTX 3090
- Official int4 and Q4_0 builds on Hugging Face and Kaggle; supported in Ollama, LM Studio, MLX and llama.cpp
Gemini 2.5 Flash
Gemini 2.5 Flash is Google's price-performance Gemini 2.5 model, released in preview with a major upgrade in reasoning over 2.0 Flash and a thinking budget that lets developers control how much the model thinks.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms, Consumer app
- Preview model gemini-2.5-flash-preview-04-17 in the Gemini API via Google AI Studio and Vertex AI
- An updated preview (gemini-2.5-flash-preview-05-20) shipped at Google I/O 2025
- Stable gemini-2.5-flash became generally available on June 17, 2025
- Also available to everyone in the Gemini app
MAI-DS-R1
MAI-DS-R1 is an open-weights version of DeepSeek R1 post-trained by the Microsoft AI team to respond to more previously blocked topics and reduce risk while keeping the original model's reasoning capabilities.
- Modalities
- text → text
- License
- MIT
- At launch
- Open weights, API
- Post-trained from DeepSeek-R1
- Released on Azure AI Foundry and Hugging Face
- Retired from Microsoft Foundry on February 27, 2026
Granite 3.3
Update of the Granite 2B and 8B models with improved reasoning that can be toggled, fill-in-the-middle support and a 128K-token context for the 8B model.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 2B and 8B
- License
- Apache-2.0
- At launch
- Open weights
- Released alongside Granite Speech 3.3 8B, an audio-in, text-out speech model
- Base models supersede their Granite 3.1 predecessors
- Released under Apache 2.0
OpenAI o3 and o4-mini
o3 and o4-mini are o-series reasoning models and the first that can agentically use and combine every ChatGPT tool, including web search, Python, file analysis, visual reasoning, and image generation.
- Modalities
- text, image → text
- Context
- 200K context
- At launch
- Consumer app, API
- API model IDs o3 and o4-mini (snapshots dated 2025-04-16)
- Replaced o1, o3-mini, and o3-mini-high for ChatGPT Plus, Pro, and Team users
- ChatGPT also offered o4-mini-high; free users could try o4-mini via 'Think'
- 200,000-token context window per the API model pages
Seedream 3.0
Seedream 3.0 is ByteDance Seed's bilingual Chinese-English text-to-image model, adding native 2K output, faster generation, and more accurate small-text rendering and layout compared with Seedream 2.0.
- Modalities
- text → image
- At launch
- Consumer app, API
- Native 2K output across aspect ratios without post-processing
- Launched in April 2025 and fully available in Doubao and Jimeng
- Volcano Engine API model ID doubao-seedream-3.0-t2i-250415
UI-TARS-1.5
UI-TARS-1.5 is a multimodal GUI agent built on a vision-language model and improved with reinforcement learning so it reasons before acting in computer-use and game tasks; ByteDance Seed open-sourced a 7B version.
- Modalities
- text, image → text
- Parameters
- 7B
- License
- Apache-2.0
- At launch
- Open weights
- UI-TARS-1.5-7B weights released on Hugging Face under Apache-2.0
- Adds game scenarios supporting up to 1,000 interactions
- The open 7B model targets general computer use rather than games
Embed 4
Multimodal embedding model for searching complex documents such as PDF reports and slides that mix text, images, tables, graphs and code.
- Modalities
- text, image → embedding
- Context
- 125K context
- At launch
- API
- Embeds documents up to 128K tokens, about 200 pages
- Supports 100+ languages and cross-lingual search
- Matryoshka embeddings for flexible storage
Kling 2.0
Kling 2.0 Master is the second-generation Kling video generation model, launched globally with improvements in motion quality, semantic responsiveness, and visual aesthetics, alongside the Kolors 2.0 image model.
- Modalities
- text, image → video
- At launch
- Consumer app
- Launched at Kling's From Vision to Screen event in Beijing
- Released together with Kolors 2.0 and new multimodal editing features
- Kuaishou reported more than 20 Kling iterations since June 2024
Kolors 2.0
Kolors 2.0 is Kuaishou's second-generation image generation model, launched on Kling AI with the Kling 2.0 video model and improved prompt adherence, cinematic quality, and artistic style expression.
- Modalities
- text, image → image
- At launch
- Consumer app
- Launched on Kling AI alongside Kling 2.0 Master
- Improves prompt adherence, cinematic quality, and artistic styles
GLM-4-0414 and GLM-Z1
GLM-4-0414 is Z.ai's open-weights model series at 9B and 32B, released with the GLM-Z1 deep reasoning models and GLM-Z1-Rumination, which thinks longer on open-ended research tasks. GLM-4-32B-Base-0414 was pre-trained on 15T tokens.
- Modalities
- text → text
- Parameters
- 9B, 32B
- License
- MIT
- At launch
- API, Open weights
- GLM-Z1-32B-0414 adds cold-start and extended reinforcement learning for math, code, and logic
- API launched GLM-4-Air-250414, GLM-4-Flash-250414, and GLM-Z1-Air, AirX, and Flash
- Weights under the MIT license
GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano
GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano are API models that improve on GPT-4o in coding and instruction following, with context windows up to 1 million tokens and a June 2024 knowledge cutoff.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API
- API model IDs gpt-4.1, gpt-4.1-mini, and gpt-4.1-nano
- 1,047,576-token context window for all three, per the API model pages
- GPT-4.1 nano is OpenAI's first nano model
- GPT-4.1 came to ChatGPT for all paid users on May 14, 2025
Nemotron-H
Family of hybrid base language models from NVIDIA that replace most attention layers with Mamba-2 layers to speed up inference, released under a research license.
- Modalities
- text → text
- License
- NVIDIA Internal Scientific Research and Development Model License
- At launch
- Open weights
- The 8B base model uses only four attention layers
- Described in the Nemotron-H technical report
- Base completion models intended for customization
Seed-Thinking-v1.5
Seed-Thinking-v1.5 is ByteDance Seed's reasoning model trained with large-scale reinforcement learning for math, coding, science, and creative writing; Volcano Engine serves it as Doubao-1.5-thinking-pro.
- Modalities
- text → text
- Parameters
- MoE: 200B total, 20B active
- At launch
- API
- Mixture-of-experts with 200B total and 20B active parameters
- API access on Volcano Engine opened April 17, 2025
- Technical report published on GitHub
Kimi-VL
Kimi-VL is an efficient open-weights mixture-of-experts vision-language model for multimodal reasoning, long-context understanding, and agent tasks. It was released as Kimi-VL-A3B-Instruct and Kimi-VL-A3B-Thinking.
- Modalities
- text, image, video → text
- Context
- 125K context
- Parameters
- MoE: 16B total, 3B active
- License
- MIT
- At launch
- Open weights
- Activates about 2.8B parameters in its language decoder
- 128K extended context window
- Native-resolution MoonViT vision encoder for high-resolution inputs
Amazon Nova Sonic
Amazon Nova Sonic is Amazon's first speech-to-speech foundation model, unifying speech understanding and generation in one model so developers can build real-time voice conversations without separate recognition and text-to-speech models.
- Modalities
- text, audio → text, audio
- Context
- 300K context
- At launch
- API
- Launched with American and British English, adding Spanish in June 2025 and French, Italian and German in July 2025
- 300K token context window
- Launched in Amazon Bedrock in US East (N. Virginia)
Amazon Nova Reel 1.1
Amazon Nova Reel 1.1 updates Amazon's video generation model to produce multi-shot videos of up to two minutes from 6-second shots, with style consistency across shots and improved single-shot quality and latency.
- Modalities
- text, image → video
- At launch
- API
- Generates videos up to 2 minutes long
- Automated mode from one prompt or manual storyboard mode with up to 20 shots
- Available in Amazon Bedrock in US East (N. Virginia)
Gen-4 Turbo
Faster and more efficient version of Gen-4, described by Runway as its fastest and most efficient video model at launch and made available to all users.
- Modalities
- text, image → video
- At launch
- Consumer app
- Available to all plans at launch
- Turbo variant of Gen-4
Llama-3.1-Nemotron-Ultra-253B-v1
Largest Llama Nemotron reasoning model, derived from Meta Llama-3.1-405B-Instruct with neural architecture search to cut memory use while keeping accuracy.
- Modalities
- text → text
- Parameters
- 253B
- License
- NVIDIA Open Model License
- At launch
- Open weights
- Post-trained for reasoning, chat preferences, RAG and tool calling
- Trained between November 2024 and April 2025
- Part of the Llama Nemotron collection
Llama 4 Maverick
Llama 4 Maverick is a natively multimodal mixture-of-experts model with 17B active parameters and 128 experts, released with Llama 4 Scout as the first Llama 4 models.
- Modalities
- text, image → text
- Context
- 1M context
- Parameters
- MoE: 400B total, 17B active, 128 experts
- License
- Llama 4 Community License
- At launch
- Open weights, Cloud platforms, Consumer app
- 400B total parameters; runs on a single NVIDIA H100 DGX host
- Model card lists a 1M-token context window and 12 supported languages
- Knowledge cutoff of August 2024
- Powers Meta AI in WhatsApp, Messenger, Instagram Direct and meta.ai
Llama 4 Scout
Llama 4 Scout is a natively multimodal mixture-of-experts model with 17B active parameters and 16 experts, notable for a 10 million token context window.
- Modalities
- text, image → text
- Context
- 10M context
- Parameters
- MoE: 109B total, 17B active, 16 experts
- License
- Llama 4 Community License
- At launch
- Open weights, Cloud platforms
- 109B total parameters; fits on a single NVIDIA H100 GPU with Int4 quantization
- Accepts multilingual text and image input
- Llama 4 Behemoth was previewed but not released because it was still training
Motion 2.0
Leonardo.Ai's own video generation model, built into its AI Creation interface to make videos from a text prompt or from images, with boosted frame rates for smoother playback.
- Modalities
- text, image → video
- At launch
- Consumer app
- Launched for paid users on web and mobile
- 720p output added on May 6, 2025
- Offered on Replicate with Phoenix 1.0 from June 10, 2025
Midjourney V7
Seventh-generation image model with better text and image prompt handling, higher image quality and coherence, personalization on by default, and a new Draft Mode that renders at ten times the speed.
- Modalities
- text, image → image
- At launch
- Consumer app
- Launched as an alpha with Turbo and Relax modes
- Became the default model on June 17, 2025
- Introduced Omni Reference and Draft Mode
MiniMax Speech-02
Speech-02 is MiniMax's text-to-audio and voice cloning model series, offered as Speech-02-HD for high-fidelity voiceovers and audiobooks and Speech-02-Turbo for low-latency interactive use.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Voice cloning from 10 seconds of recorded audio
- Library of more than 300 prebuilt voices
- Built on an autoregressive Transformer with a learnable speaker encoder, per MiniMax's May 2025 technical post
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 April 2025." The AI Engineering, updated Apr 30, 2025. https://theaiengineering.com/releases/2025/04
HTML link
<a href="https://theaiengineering.com/releases/2025/04">AI models released in April 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