AI models released in June 2026
35 releases from 20 labs in June 2026, including Claude Sonnet 5, LongCat-2.0 and Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image). Every entry is dated to the lab's official announcement and links to it.
- Releases
- 35
- Labs
- 20
- Open weights
- 13
June 2026
35 releasesClaude Sonnet 5
Claude Sonnet 5 is the next generation of Anthropic's Sonnet line, built for agentic work such as planning, browser and terminal tool use, and autonomous runs, with performance Anthropic describes as close to Claude Opus 4.8.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms, Enterprise
- API model ID claude-sonnet-5; 1M-token context window and 128K max output tokens
- Default model for Claude Free and Pro plans
- Adaptive thinking on by default; manual extended thinking returns an error
- Updated tokenizer can map the same input to roughly 1.0 to 1.35 times as many tokens
LongCat-2.0
LongCat-2.0 is Meituan's large mixture-of-experts language model for coding and agentic work, trained entirely on AI ASIC superpods and introducing LongCat Sparse Attention for efficient long-context processing.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.6T total, about 48B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- 1.6 trillion total parameters with about 48B activated per token
- Trained on 1M-context data using LongCat Sparse Attention
- Pretrained on more than 35 trillion tokens
- Weights released under MIT; API billing launched the same day
Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image)
Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image) is the fastest and most cost-efficient model in the Nano Banana family, built for high-throughput image generation and editing.
- Modalities
- text, image, video → text, image
- Context
- 64K context
- At launch
- API, Cloud platforms, Consumer app
- Generally available at launch in Google AI Studio, the Gemini API and Gemini Enterprise Agent Platform
- Recommended replacement for the original Nano Banana (gemini-2.5-flash-image)
- Also rolled out in AI Mode in Search and the Gemini app
LFM2.5-230M
Liquid AI's smallest model at release, a 230M-parameter LFM2-architecture model pre-trained on 19T tokens with a 32K context extension and a light post-training recipe meant as a base for fine-tuning in agentic workflows.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 230M
- License
- LFM Open License v1.0
- At launch
- Open weights
- Released as LFM2.5-230M-Base and post-trained LFM2.5-230M
- Post-training: distillation from LFM2.5-350M, DPO, and multi-domain RL
- Demonstrated as a skill-selection layer on a Unitree G1 humanoid robot
- Available on Hugging Face
Mistral OCR 4
OCR model that returns bounding boxes, typed block classification and inline confidence scores with extracted text, supporting 170 languages and single-container self-hosting.
- Modalities
- text, image → text
- At launch
- API, Enterprise
- Ingestion component of Mistral's Search Toolkit
- OCR 4.1 followed on July 16, 2026 with block-level confidence scores
- API model ID mistral-ocr-4-0
Seed2.1
Seed2.1 is ByteDance Seed's model family for agentic productivity, offered in Pro and Turbo sizes, with more reliable multi-step agent execution, end-to-end coding delivery, and stronger understanding of complex documents and video than Seed2.0.
- Modalities
- text, image, video → text
- At launch
- Consumer app, API
- Released in Pro and Turbo sizes
- Available to Doubao app and Volcano Engine users at launch
- Volcano Engine's June 2026 release notes also add doubao-seed-evolving, a Seed model updated at least weekly
Unlimited-OCR
Unlimited-OCR is an open vision-language OCR model from Baidu aimed at one-shot, long-horizon parsing of documents, which Baidu describes as pushing the DeepSeek-OCR approach one step further.
- Modalities
- image, text → text
- License
- MIT
- At launch
- Open weights, API
- Weights on Hugging Face and ModelScope under the MIT license
- vLLM inference support added June 28, 2026
- Made available on Baidu Cloud on July 3, 2026
LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M
Two multilingual retrieval models, the first bidirectional members of the LFM family: a single-vector embedding model and a late-interaction ColBERT model, both 350M parameters and built from LFM2.5-350M-Base.
- Modalities
- text → embedding
- Parameters
- 350M each
- License
- LFM Open License v1.0
- At launch
- Open weights
- Supports 11 languages for multilingual and cross-lingual search
- Causal attention and short convolutions converted to bidirectional versions
- Positioned for short-context search; 512-token document length per model card
- Available on Hugging Face
Kling 3.0 Turbo
Kling 3.0 Turbo is a new model in Kuaishou's Kling 3.0 video series, launched as part of a series upgrade that also improved video editing in Kling 3.0 Omni.
- Modalities
- text, image → video
- At launch
- Consumer app
- Listed as a Kling 3.0 model series upgrade in Kling's release notes
- Launched alongside upgraded video editing in Kling 3.0 Omni
GLM-5.2
GLM-5.2 is Z.ai's flagship model for long-horizon tasks and the first GLM model to deliver its capabilities on a 1M-token context. It adds multiple thinking effort levels for balancing coding performance and latency.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 744B total, 40B active
- License
- MIT
- At launch
- API, Open weights
- IndexShare reuses one indexer across every four sparse attention layers, cutting per-token FLOPs at 1M context
- Improved multi-token prediction layer for speculative decoding
- Up to 128K output tokens
- Weights under the MIT license
Kimi K2.7 Code
Kimi K2.7 Code is a coding-focused agentic model built on Kimi K2.6, with improvements on real-world long-horizon coding tasks. Moonshot says it uses about 30% fewer thinking tokens than K2.6.
- Modalities
- text, image → text
- Context
- 250K context
- Parameters
- MoE: 1T total, 32B active
- License
- Modified MIT License
- At launch
- API, Open weights
- Released through the Kimi Open Platform API, followed by a HighSpeed variant
- MoonViT vision encoder with 400M parameters
- Weights on Hugging Face under the Modified MIT License
Claude Fable 5
Claude Fable 5 is a Mythos-class Anthropic model released with safeguards for general use, aimed at long, complex software engineering, knowledge work, and research tasks. Requests on some sensitive topics are answered by Claude Opus 4.8 instead.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms, Enterprise
- API model ID claude-fable-5; 1M-token context window and 128K max output tokens
- Claude Mythos 5, the same model with some safeguards lifted, launched only for Project Glasswing partners
- Access was suspended June 12, 2026 under a US export control directive and restored July 1, 2026
- Adaptive thinking is the only thinking mode
Gemini 3.5 Live Translate
Gemini 3.5 Live Translate is an audio model for near real-time speech-to-speech translation that detects more than 70 languages and preserves the speaker's intonation, pacing and pitch.
- Modalities
- audio → audio, text
- Context
- 128K context
- At launch
- API, Enterprise
- Public preview for developers in the Gemini Live API and Google AI Studio
- Private preview for select Google Workspace business customers in Google Meet
- Generates translated speech continuously instead of waiting for the speaker to finish
North Mini Code
Cohere's first agentic coding model, a mixture-of-experts model trained specifically for agentic coding whose small active parameter footprint lets it run on local hardware.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 30B total, 3B active
- License
- Apache-2.0
- At launch
- API, Open weights
- 256K input and 64K output tokens
- Available through the Chat V2 API and as open weights
- Supported in Cohere's Model Vault
Ray3.2
Ray3 update designed with entertainment, advertising and gaming creatives that adds frame-level direction and, for the first time, exposes the full Ray control surface through an API.
- Modalities
- text, image, video → video
- At launch
- Consumer app, API
- Multi-keyframe direction with up to 16 keyframes
- Native HDR generation with EXR output
- Modify Video V2 up to 20 seconds at 1080p
SCAIL-2
SCAIL-2 is Z.ai's open-weights character animation model, which unifies controlled character animation with replacement in one video generation framework. It succeeds the SCAIL-Preview research release.
- Modalities
- text, image, video → video
- License
- MIT
- At launch
- Open weights
- Model and inference code released on June 9, 2026
- Multi-reference inference added on June 13, 2026
- Weights on Hugging Face under the MIT license
ADM 3 Cloud
ADM 3 Cloud is Apple's third-generation server image model for generating and editing images, powering photo editing tools, the new Image Playground and Genmoji.
- Modalities
- text, image → image
- At launch
- Consumer app
- Runs on Private Cloud Compute
- Specialized adapters support features such as Spatial Reframing in Photos
AFM 3 Cloud
AFM 3 Cloud is the server-side workhorse of Apple's third-generation foundation models, optimized for speed and efficiency with multimodal reasoning on Private Cloud Compute.
- Modalities
- text, image → text
- At launch
- Consumer app
- Runs on Private Cloud Compute on Apple silicon servers
- Apple says user data is never stored or shared, including with Apple
AFM 3 Cloud Pro
AFM 3 Cloud Pro is Apple's most capable server-based foundation model in the third generation, powering demanding Apple Intelligence uses such as agentic tool use and complex reasoning.
- Modalities
- text → text
- At launch
- Consumer app
- Runs on Private Cloud Compute extended to NVIDIA GPUs in Google Cloud
- Part of a model family custom-built in collaboration with Google
AFM 3 Core
AFM 3 Core is the third-generation version of Apple's dense 3B on-device foundation model, a step up in quality optimized for Apple silicon and part of a family built with Google.
- Modalities
- text, image → text
- Parameters
- 3B (dense)
- At launch
- Consumer app, API
- Successor to Apple's 3B dense on-device model
- Image understanding evaluated alongside text
- Third-generation family custom-built in collaboration with Google
AFM 3 Core Advanced
AFM 3 Core Advanced is Apple's most powerful on-device model, a natively multimodal sparse model that supports expressive voices and higher-accuracy dictation on Apple's most capable devices.
- Modalities
- text, audio → text, audio
- Parameters
- Sparse: 20B total, 1B to 4B active
- At launch
- Consumer app
- Activates 1 to 4 billion of its 20 billion parameters depending on the request
- Keeps full weights in flash memory and loads active experts into DRAM as needed
- Limited to Apple's most capable Apple silicon systems
Magenta RealTime 2
Magenta RealTime 2 is an open-weights live music model from Google DeepMind's Magenta team for real-time, interactive generation on a laptop. It follows Magenta RealTime with frame-wise generation, MIDI control and much lower control latency.
- Modalities
- text, audio → audio
- Parameters
- 2.4B (base), 230M (small)
- License
- CC-BY-4.0
- At launch
- Open weights
- Base model has 2.4B parameters and small model 230M
- Control latency of about 200 ms versus about 3 seconds for the original
- MLX-based C++ engine runs natively on Apple Silicon
- Weights under CC BY 4.0, code under Apache 2.0
Nemotron 3 Ultra
Frontier-scale Nemotron 3 model for complex agentic workflows, long-context analysis and multilingual reasoning, built on a hybrid latent mixture-of-experts design.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 550B total, 55B active
- License
- OpenMDW-1.1
- At launch
- Open weights
- Supports up to 1M tokens of context
- Includes multi-token prediction layers
- Released under the OpenMDW 1.1 license
Gemma 4 12B
Gemma 4 12B is a mid-sized, encoder-free multimodal Gemma 4 model in which vision and audio inputs flow directly into the language model backbone. It sits between the E4B and 26B MoE models and targets laptops.
- Modalities
- text, image, audio → text
- Context
- 256K context
- Parameters
- 12B
- License
- Apache-2.0
- At launch
- Open weights, Cloud platforms
- No separate vision or audio encoders; inputs are projected into the LLM embedding space
- First mid-sized Gemma model with native audio input
- Runs locally with 16GB of VRAM or unified memory
- 256K-token context per the Hugging Face model card; ships with Multi-Token Prediction drafters
GPT-Rosalind-5.5
GPT-Rosalind-5.5 updates OpenAI's GPT-Rosalind life sciences series, combining GPT-5.5's agentic coding and tool-use capabilities with stronger model intelligence in drug-discovery domains such as medicinal chemistry and genomics.
- At launch
- Enterprise
- With this release, OpenAI expanded access to the GPT-Rosalind series to eligible organizations globally through trusted access
- Added Codex plugins for evidence retrieval, biological interpretation, and bioinformatics, open to all users and powered by GPT-Rosalind for qualified enterprise users
- Added interactive Codex viewers for sequence, alignment, and structure files
Grok Imagine Video 1.5
Grok Imagine Video 1.5 is SpaceXSI's image-to-video model with generated audio. It launched in API preview, then became generally available with improved motion, physics, and synced sound and speech.
- Modalities
- text, image → video, audio
- At launch
- API
- Preview model ID grok-imagine-video-1.5-preview, with clips up to 720p
- GA as grok-imagine-video-1.5 on June 16, 2026; Video 1.5 Fast added to grok.com and apps
- July 31, 2026 update added image and voice references, text-to-video, and native 1080p
Ideogram 4.0
Ideogram's first open-weight foundation model, a 9.3B-parameter single-stream diffusion transformer trained from scratch with a vision-language text encoder and structured JSON prompts.
- Modalities
- text → image
- Parameters
- 9.3B
- License
- Ideogram 4 Non-Commercial License
- At launch
- Open weights, Consumer app, API
- 9.3B-parameter DiT trained from scratch
- Weights on Hugging Face, code on GitHub
- Production use requires a commercial license
MAI-Code-1-Flash
MAI-Code-1-Flash is Microsoft AI's inference-efficient agentic coding model, trained inside the agent tooling GitHub Copilot uses and rolled out to GitHub Copilot users in Visual Studio Code. It uses adaptive thinking to scale reasoning to task difficulty.
- Modalities
- text → text
- Parameters
- 5B active
- At launch
- Consumer app
- Rolled out to GitHub Copilot individual users in VS Code
- Has 5 billion active parameters
- Trained without distillation from third-party models
MAI-Image-2.5
MAI-Image-2.5 is Microsoft AI's image model for high-quality text-to-image generation and precise, localized image editing, with face and identity consistency across edits. It launched in Foundry and in PowerPoint.
- Modalities
- text, image → image
- At launch
- API, Enterprise
- Available to developers in Microsoft Foundry
- Live in PowerPoint and rolling out to OneDrive for photo editing
- Launched with a faster MAI-Image-2.5-Flash variant
MAI-Image-2.5-Flash
MAI-Image-2.5-Flash is a faster, lower-cost variant of Microsoft AI's MAI-Image-2.5 for scalable production image generation and editing workloads, launched in Microsoft Foundry on the same day as MAI-Image-2.5.
- Modalities
- text, image → image
- At launch
- API
- Supports both text-to-image generation and image editing
- Available in Microsoft Foundry
- Part of the seven-model MAI launch at Microsoft Build 2026
MAI-Thinking-1
MAI-Thinking-1 is Microsoft AI's flagship reasoning model, a medium-sized sparse mixture-of-experts model trained from scratch without distillation from third-party models, aimed at software engineering, math and enterprise tasks.
- Modalities
- text → text
- Parameters
- MoE: about 1T total, 35B active
- At launch
- API
- One of seven in-house models Microsoft AI announced on June 2, 2026
- Public preview in Microsoft Foundry from August 12, 2026
- Sparse mixture-of-experts with about 1T total and 35B active parameters
MAI-Transcribe-1.5
MAI-Transcribe-1.5 is Microsoft AI's multilingual speech-to-text model, expanding language coverage from 25 to 43 languages and adding keyword biasing for domain-specific terminology, and is being integrated into Copilot, Teams, GitHub and Dynamics 365.
- Modalities
- audio → text
- At launch
- API, Enterprise
- Covers 43 languages, up from 25 in MAI-Transcribe-1
- Adds keyword biasing for user-supplied domain terms
- Available in Microsoft Foundry
MAI-Voice-2
MAI-Voice-2 is Microsoft AI's text-to-speech model, expanding from English-only to 15 languages with emotion tags and zero-shot voice prompting from short reference audio, protected by consent guardrails.
- Modalities
- text, audio → audio
- At launch
- API
- Supports 15 languages
- Zero-shot voice prompting from 5 to 60 seconds of reference audio
- Available in Microsoft Foundry and being integrated into VS Code and Dynamics 365 Contact Center
MiniMax M3
MiniMax M3 is a natively multimodal open-weight model aimed at coding and agentic work. It introduces MiniMax Sparse Attention (MSA) to reach a context window of up to 1M tokens and accepts image and video input.
- Modalities
- text, image, video → text
- Context
- 1M context
- Parameters
- MoE: about 428B total, 23B active
- License
- MiniMax Community License
- At launch
- API, Open weights
- About 428B total parameters with about 23B activated, per the model card
- Supports three reasoning modes through a thinking parameter
- Can operate a desktop computer, enabling computer use in MiniMax Code
- MSA is open-sourced as a separate GitHub repository
Qwen3.7-Plus
Qwen3.7-Plus is a multimodal agent model that unifies vision and language, building on the Qwen3.7 text backbone. It can read screens, operate GUIs, write code from visual references, and navigate mobile apps.
- Modalities
- text, image, video → text
- At launch
- API
- Accepts text, image, and video input
- Supports preserving thinking content across turns for agentic tasks
- Available through Alibaba Cloud Model Studio
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 2026." The AI Engineering, updated Jun 30, 2026. https://theaiengineering.com/releases/2026/06
HTML link
<a href="https://theaiengineering.com/releases/2026/06">AI models released in June 2026</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering