AI models released in June 2025
32 releases from 17 labs in June 2025, including ERNIE 4.5 (open-weight family), Seed1.6-Embedding and Hunyuan-A13B. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 32
- Labs
- 17
- Open weights
- 15
June 2025
32 releasesERNIE 4.5 (open-weight family)
Baidu open-sourced the ERNIE 4.5 family of 10 variants: mixture-of-experts language and vision-language models with 47B or 3B active parameters plus a 0.3B dense model, trained with PaddlePaddle and offered in PyTorch-compatible formats.
- Modalities
- text, image → text
- Context
- 128K context
- Parameters
- MoE: up to 424B total, 47B active; plus a 0.3B dense model
- License
- Apache-2.0
- At launch
- Open weights
- Largest models are ERNIE-4.5-300B-A47B (text) and ERNIE-4.5-VL-424B-A47B (vision-language)
- 131,072-token context length per the Hugging Face model cards
- Vision-language models support thinking and non-thinking modes
- Released with the ERNIEKit fine-tuning and FastDeploy deployment toolkits
Seed1.6-Embedding
Seed1.6-Embedding is a multimodal embedding model built on Seed1.6-Flash that supports hybrid retrieval across text, image, and video, with selectable 2048 or 1024 embedding dimensions.
- Modalities
- text, image, video → text
- At launch
- API
- API model ID doubao-embedding-vision-250615 on Volcano Engine
- Dual-tower design using the final-layer EOS token representation
- Supports 2048 and 1024 embedding dimensions
Hunyuan-A13B
Hunyuan-A13B is an open fine-grained mixture-of-experts language model with 80B total and 13B active parameters, supporting hybrid fast and slow thinking for reasoning and agent tasks in resource-constrained environments.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 80B total, 13B active
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Native 256K context window
- Pretrain, Instruct, FP8, and GPTQ-Int4 checkpoints released with a technical report
- Defaults to slow-thinking mode, which can be turned off
Qwen-TTS
Qwen-TTS is Qwen's text-to-speech model on the Qwen API, trained on millions of hours of speech. This update added support for several Chinese dialects.
- Modalities
- text → audio
- At launch
- API
- Model IDs qwen-tts-latest and qwen-tts-2025-05-22
- Generates 3 Chinese dialects: Pekingese, Shanghainese, and Sichuanese
- Available through the Qwen API
FLUX.1 Kontext [dev]
Open-weight developer version of FLUX.1 Kontext focused exclusively on image editing, a 12B-parameter model that runs on consumer hardware and supports iterative, local and global edits.
- Modalities
- text, image → image
- Parameters
- 12B
- License
- FLUX.1 [dev] Non-Commercial License
- At launch
- Open weights, Cloud platforms
- 12B parameters, compatible with FLUX.1 [dev] inference code
- BF16, FP8 and FP4 TensorRT variants for NVIDIA Blackwell
- Released alongside the KontextBench evaluation set and a technical report
Keye-VL-8B-Preview
Keye-VL is Kuaishou's first open multimodal large language model, built to understand short-form video as well as images, with a switchable thinking mode for reasoning.
- Modalities
- text, image, video → text
- Parameters
- 8B
- License
- Apache-2.0
- At launch
- Open weights
- Focused on dynamic, information-dense short-form video
- Soft switch between thinking and non-thinking modes
- Technical report followed on July 1, 2025
o3-deep-research and o4-mini-deep-research
o3-deep-research and o4-mini-deep-research bring the deep research models used in ChatGPT to the API, letting developers trigger multi-step research that can pull external context through web search and MCP servers.
- Modalities
- text, image → text
- Context
- 200K context
- At launch
- API
- Snapshots o3-deep-research-2025-06-26 and o4-mini-deep-research-2025-06-26
- Launched with API webhooks; OpenAI recommends background mode for long tasks
- 200,000-token context window per the API model pages
Qwen VLo
Qwen VLo is a preview unified multimodal understanding and generation model from Qwen, which generates high-quality image recreations based on its understanding of the input.
- Modalities
- text, image → image
- At launch
- Consumer app
- Preview version available through Qwen Chat
- Supports instructions in multiple languages, including Chinese and English
- Extends Qwen's vision models from understanding the world to depicting it
Seed1.6
Seed1.6 is ByteDance Seed's general-purpose model series combining multimodal understanding, adaptive deep thinking, GUI interaction, and a 256K context window, offered on Volcano Engine with Seed1.6-Thinking and Seed1.6-Flash variants.
- Modalities
- text, image → text
- Context
- 250K context
- Parameters
- MoE: 230B total, 23B active
- At launch
- API
- Sparse MoE pretrained with 23B active and 230B total parameters
- Adaptive chain-of-thought (AdaCoT) decides when to think based on difficulty
- Long-context training raised sequence length from 32K to 256K
- Listed in Volcano Engine's June 2025 model release notes as doubao-seed-1-6-250615 with thinking and flash variants
Voice Design v3
Text-to-voice model that creates custom voices from a written description, returning three options per prompt for immediate use in projects.
- Modalities
- text → audio
- At launch
- Consumer app, API
- Model ID eleven_ttv_v3
- Two design modes
- Generates three voice options per description
Kimi-VL-A3B-Thinking-2506
Kimi-VL-A3B-Thinking-2506 is an improved version of Moonshot AI's Kimi-VL-A3B-Thinking vision-language reasoning model. Moonshot recommends using it instead of the earlier Thinking checkpoint released in April 2025.
- Modalities
- text, image, video → text
- Context
- 125K context
- Parameters
- MoE: 16B total, 3B active
- License
- MIT
- At launch
- Open weights
- Released with a tech blog and cookbook on Hugging Face
- Same 16B-total, 3B-active mixture-of-experts design as Kimi-VL
- Weights under the MIT License
Magenta RealTime
Magenta RealTime is an open-weights live music model from Google's Magenta team that generates music in two-second chunks steered by text or audio prompts. It is the open counterpart to the Lyria RealTime API.
- Modalities
- text, audio → audio
- Parameters
- 800M
- License
- CC-BY-4.0
- At launch
- Open weights
- 800M-parameter autoregressive transformer
- Released as a research preview
- Weights on Hugging Face and Google Cloud Storage; weights under CC BY 4.0, code under Apache 2.0
MiniMax Music 1.5
MiniMax Music 1.5 is a music generation model that creates songs from style prompts and lyrics, extending generation to full four-minute tracks with more natural vocals, layered instrumentals, and coherent song structure.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Beta listed in the API release notes on June 20, 2025
- Full release with 4-minute songs announced September 11, 2025
- Advanced mode assigns lyrics to sections such as Intro, Verse, and Chorus
Mistral Small 3.2
Minor update of Mistral Small 3.1 with a more robust function calling template, released as open weights and added to the API three days later.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- API model ID mistral-small-2506, available June 23, 2025
- Accepts text and image input
- Deprecated April 30, 2026 in favor of Mistral Small 4
Midjourney Video V1
Midjourney's first video model, an image-to-video workflow that animates Midjourney images with automatic or manual motion prompts and low or high motion settings.
- Modalities
- image, text → video
- At launch
- Consumer app
- Each job produces four 5-second videos
- Videos can be extended about 4 seconds at a time, four times
- Launched web-only; HD 720p mode added August 2025
MiniMax Hailuo 02
MiniMax Hailuo 02 is a video generation model with native 1080p output, stronger instruction following, and better rendering of complex physics, built on a new architecture with three times the parameters of its predecessor.
- Modalities
- text, image → video
- At launch
- API, Consumer app
- Offered in 768p-6s, 768p-10s, and 1080p-6s versions
- Trained on four times the data of the previous model
- Available on Hailuo Video web and mobile and the MiniMax API
Gemini 2.5 Flash-Lite
Gemini 2.5 Flash-Lite is the most cost-efficient and fastest model in the Gemini 2.5 family, aimed at high-volume, latency-sensitive tasks such as translation and classification.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms
- Preview released June 17, 2025, the same day 2.5 Pro and 2.5 Flash became generally available
- Stable gemini-2.5-flash-lite became generally available on July 22, 2025
- 1 million token context window, controllable thinking budgets and native tools such as Grounding with Google Search
Kimi-Dev-72B
Kimi-Dev-72B is Moonshot AI's open-weights coding model for software engineering issue resolution. It was optimized with large-scale reinforcement learning in which it patches real repositories in Docker and is rewarded only when the full test suite passes.
- Modalities
- text → text
- Parameters
- 72B
- License
- MIT
- At launch
- Open weights
- Targets SWE-bench-style repository issue resolution
- Available on Hugging Face and GitHub
- Weights under the MIT License
MiniMax-M1
MiniMax-M1 is an open-weight hybrid-attention reasoning model built on MiniMax-Text-01, combining Lightning Attention with mixture-of-experts and trained with large-scale reinforcement learning using MiniMax's CISPO algorithm.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 456B total, 45.9B active
- License
- Apache-2.0
- At launch
- API, Consumer app, Open weights
- Native 1M-token context with up to 80K tokens of reasoning output
- Released in 40K and 80K thinking-budget versions
- Weights under the Apache-2.0 license
- Weights and technical report published on Hugging Face and GitHub
Hunyuan3D 2.1
Hunyuan3D 2.1 is a fully open 3D asset generation system that releases full model weights and training code, replacing the earlier RGB texture model with a physically based rendering (PBR) material pipeline.
- Modalities
- image → 3d
- License
- Tencent Hunyuan 3D 2.1 Community License Agreement
- At launch
- Open weights
- Includes image-to-shape and PBR texture generation models
- Technical report published June 19, 2025
Granite Vision 3.3
Compact IBM vision-language model for visual document understanding that automates content extraction from tables, charts, infographics, plots and diagrams, succeeding Granite Vision 3.2.
- Modalities
- text, image → text
- Parameters
- 2B
- License
- Apache-2.0
- At launch
- Open weights
- Released with a Granite Vision 3.3 2B embedding variant
- Successor to Granite Vision 3.2
- Released under Apache 2.0
Seedance 1.0
Seedance 1.0 is ByteDance Seed's foundation video generation model, producing 1080p video from text and image prompts with native multi-shot storytelling, stable motion, and fast inference.
- Modalities
- text, image → video
- At launch
- Consumer app, API
- Generates 10-second clips with 2 to 3 shot transitions
- Available through Jimeng, Doubao, and the Volcano Engine API
- Volcano Engine offers Seedance 1.0 Pro and Lite versions
- Technical report published at launch
Magistral Medium
Enterprise version of Magistral, Mistral AI's first reasoning model, focused on domain-specific, transparent and multilingual reasoning for use across professional domains.
- Modalities
- text → text
- Context
- 40K context
- At launch
- API, Consumer app
- Released alongside the open Magistral Small
- Previewed in Le Chat
- API model ID magistral-medium-2506
Magistral Small
Open-weight version of Magistral, Mistral AI's first reasoning model, built for transparent, multilingual step-by-step reasoning that users can follow and verify.
- Modalities
- text → text
- Context
- 40K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Released alongside the enterprise Magistral Medium
- Weights published as Magistral-Small-2506
- API model ID magistral-small-2506
OpenAI o3-pro
o3-pro is a version of OpenAI o3 designed to think longer and give more reliable answers. It replaced o1-pro for ChatGPT Pro and Team users and launched in the API the same day.
- Modalities
- text, image → text
- Context
- 200K context
- At launch
- Consumer app, API
- API model ID o3-pro (snapshot o3-pro-2025-06-10)
- Has ChatGPT tools including web search, file analysis, Python, and memory
- Enterprise and Edu ChatGPT access followed the week after launch
- 200,000-token context window per the API model page
Apple Foundation Models (2025)
Apple's 2025 foundation models update an on-device model of about 3B parameters and a server model built on a new parallel track mixture-of-experts architecture, adding image understanding and 15 languages.
- Modalities
- text, image → text
- Parameters
- About 3B (on-device); server size not disclosed
- At launch
- Consumer app, API
- Foundation Models framework gives developers direct access to the on-device model
- Framework supports guided generation via Swift @Generable and tool calling
- Server model runs on Private Cloud Compute
- Vision encoders: 300M ViTDet-L on device and 1B ViT-g on the server
SeedEdit 3.0
SeedEdit 3.0 is an instruction-based image editing model built on Seedream 3.0 that better preserves subjects, backgrounds, and details, suited to portrait retouching, background changes, perspective shifts, and lighting adjustments.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Processes and outputs images above 1K resolution
- Opened for testing on the Jimeng web app at launch
- Volcano Engine later added it as doubao-seededit-3.0-i2i
Qwen3-Embedding
Qwen3-Embedding is a series of text embedding models built on Qwen3 foundation models, released in 0.6B, 4B, and 8B sizes for retrieval, clustering, and classification.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 0.6B, 4B, 8B
- License
- Apache-2.0
- At launch
- Open weights
- Released together with the Qwen3-Reranker series
- 32K context length
- Weights on Hugging Face and ModelScope under the Apache 2.0 license
Qwen3-Reranker
Qwen3-Reranker is a series of text reranking models built on Qwen3 foundation models, released alongside Qwen3-Embedding for retrieval pipelines that rerank candidates found by dense embedding search.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 0.6B, 4B, 8B
- License
- Apache-2.0
- At launch
- Open weights
- Available in 0.6B, 4B, and 8B sizes
- 32K context length
- Weights under the Apache 2.0 license
Eleven v3
Expressive text-to-speech model released as a research preview alpha, supporting 70+ languages, multi-speaker dialogue and inline audio tags such as [excited], [whispers] and [sighs].
- Modalities
- text → audio
- At launch
- Consumer app, API
- Added to the API on August 20, 2025
- Generally available from February 2, 2026
- Dialogue mode with multiple speakers
Llama Nemotron Nano VL 8B
Document intelligence vision-language model that queries and summarizes images, deployable in the data center, in the cloud and at the edge.
- Modalities
- text, image → text
- Parameters
- 8B
- License
- NVIDIA Open Model License
- At launch
- Open weights
- Focused on document understanding
- Part of the Llama Nemotron family
- Released under the NVIDIA Open Model License
Lucid Realism
Leonardo.Ai image model for cinematic, story-driven visuals, trained with image-to-video generation in mind and suited to start frames for Motion 2.0 video.
- Modalities
- text → image
- At launch
- Consumer app
- Aimed at photo-like fashion shoots, interiors, lifestyle and ads
- Built to pair with Motion 2.0 video workflows
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 June 2025." The AI Engineering, updated Jun 30, 2025. https://theaiengineering.com/releases/2025/06
HTML link
<a href="https://theaiengineering.com/releases/2025/06">AI models released in June 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