AI models released in July 2026

34 releases from 19 labs in July 2026, including DeepSeek-V4-Flash (0731), MiniMax H3 and Seedance 2.5. Every entry is dated to the lab's official announcement and links to it.

Releases
34
Labs
19
Open weights
13

July 2026

34 releases
  1. DeepSeekLLMOpen weights

    DeepSeek-V4-Flash (0731)

    DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the April preview with substantially enhanced agentic capabilities. It keeps the preview's architecture and size and was re-post-trained, with a speculative decoding module attached.

    Modalities
    text → text
    Context
    1M context
    Parameters
    MoE: 284B total, 13B active
    License
    MIT
    At launch
    API, Open weights
    • Released to the deepseek-v4-flash API as a public beta; the app, web, and V4-Pro API were unchanged
    • Supports low, high, and max reasoning effort levels
    • Weights on Hugging Face as DeepSeek-V4-Flash-0731 under the MIT License
    Source: Change Log | DeepSeek API Docs
  2. MiniMaxVideoOpen weights

    MiniMax H3

    MiniMax H3 is a general-purpose multimodal generation model that reads combined text, image, video, and audio context and generates video with native stereo sound, up to 15 seconds long at up to 2K resolution.

    Modalities
    text, image, video, audio → video, audio
    Parameters
    33B dense Omni Transformer
    License
    MiniMax H3 Community License Agreement
    At launch
    API, Consumer app, Open weights
    • Weights open-sourced on August 3, 2026 as first-and-last-frame and omni-reference checkpoints
    • Core H3-Omni-Transformer is a 33B-parameter dense single-stream Transformer
    • Outputs 4 to 15 second clips at 24 fps with 32 kHz stereo audio
    • API variants include H3 Max, H3 Highspeed, and H3 Lite
    Source: MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities
  3. ByteDance SeedVideoProprietary

    Seedance 2.5

    Seedance 2.5 is ByteDance Seed's video creation model built on the Seedance 2.0 joint audio-video architecture, extending single-pass generation to 30 seconds with multi-round extension and upgraded multimodal referencing and editing.

    Modalities
    text, image, video, audio → video, audio
    At launch
    Consumer app, API
    • Generates up to 30 seconds of audio and video per pass, up from 15 seconds
    • Accepts up to 30 images, 10 video clips, and 10 audio clips as references
    • Rolling out on Jimeng AI and Doubao; BytePlus offers it as Dreamina Seedance 2.5
    • Supports timestamp-based editing, per Volcano Engine Ark release notes
    Source: One-take Creation, Flexible Referencing: Introducing Seedance 2.5
  4. GoogleMultimodalProprietary

    Gemini Robotics ER 2

    Gemini Robotics ER 2 is Google DeepMind's most capable embodied reasoning model, acting as a high-level brain that lets robots understand the physical world, plan multi-step tasks, track progress from video and coordinate with other robots.

    Modalities
    text, image, audio, video → text
    Context
    128K context
    At launch
    API, Cloud platforms
    • Publicly available in the Gemini API and Google AI Studio, and in private preview on Gemini Enterprise Agent Platform
    • Ships as gemini-robotics-er-2-preview plus a streaming variant for the Live API
    • Supports function calling with blocking behavior for physical robot actions
    Source: Gemini Robotics ER 2
  5. Thinking Machines LabMultimodalOpen weights

    Inkling-Small

    A smaller open-weights sibling of Inkling that Thinking Machines Lab says reaches comparable performance at a quarter of the size, with native reasoning over text, images and audio.

    Modalities
    text, image, audio → text
    Context
    1M context
    Parameters
    MoE: 276B total, 12B active
    License
    Apache 2.0
    At launch
    API, Open weights
    • 276B total parameters, 12B active; 42 layers routing each token to 6 of 256 experts plus 2 shared experts
    • Context window up to 1M tokens
    • Ahead of Inkling on reasoning and agentic coding in the lab's results; Inkling kept the edge on knowledge and factuality
    • Weights on Hugging Face (original and NVFP4) under Apache 2.0; fine-tuning on Tinker and chat in the Tinker Playground
    Source: Introducing Inkling-Small
  6. xAISpeechProprietary

    Grok Voice 2.0

    Grok Voice 2.0 is SpaceXSI's next-generation speech-to-speech model, improving speech reasoning, transcription accuracy, conversational dynamics, and tool use over Grok Voice 1.0 while using reasoning tokens more efficiently.

    Modalities
    audio, text → audio
    At launch
    API
    • API model ID: grok-voice-2.0, in the Speech to Speech API
    • Set to become the grok-voice-latest target on August 5, 2026
    • Reasons through queries while speaking, with no added latency per SpaceXSI
    • Deprecated on October 14, 2026 in favor of Grok Voice 3.0
    Source: Introducing Grok Voice 2.0 | SpaceXSI
  7. GoogleMusicProprietary

    Lyria 3.5

    Lyria 3.5 is a Google DeepMind music generation model that launched first in Google Flow Music. It improves musicality, lyrics and vocal quality over Lyria 3 and adds easier control over tempo and track duration.

    Modalities
    text → audio
    At launch
    Consumer app, API
    • Rolled out first in Google Flow Music on July 29, 2026
    • More natural melodies, stronger lyric adherence and more expressive vocals
    • Generally available in the Gemini API as lyria-3.5 from September 3, 2026
    • Available in the Gemini app globally from September 4, 2026
    Source: We're launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control.
  8. OpenAISpeechProprietary

    GPT Transcribe and GPT Live Transcribe

    GPT Transcribe and GPT Live Transcribe are speech-to-text models for the API: GPT Transcribe handles audio files and committed Realtime turns, while GPT Live Transcribe provides low-latency streaming transcription with tunable latency.

    Modalities
    audio, text → text
    At launch
    API
    • API model IDs gpt-transcribe (v1/audio/transcriptions) and gpt-live-transcribe (v1/realtime)
    • Both support free-form transcription context, keyword hints, and multiple expected input languages
    • Named as migration targets for whisper-1 and the gpt-4o transcribe models, which shut down February 26, 2027
    Source: Changelog | OpenAI API
  9. Liquid AIEmbeddingOpen weights

    LFM2.5-Encoder-230M and LFM2.5-Encoder-350M

    Two bidirectional masked-language encoders on the LFM2 hybrid architecture, initialized from LFM2.5-230M and LFM2.5-350M and meant to be fine-tuned for classification, natural language understanding, token-level tasks, and retrieval.

    Modalities
    text → embedding
    Context
    8K context
    Parameters
    230M; 350M
    License
    LFM Open License v1.0
    At launch
    Open weights
    • Native 8,192-token context length
    • Trained with a masked language modeling objective that masks 30% of tokens
    • Demo fine-tunes include PII detection, spell checking, and a policy linter
    • Available on Hugging Face
    Source: LFM2.5-Encoders: Fast at Long Context, Even on CPU
  10. AnthropicLLMProprietary

    Claude Opus 5

    Claude Opus 5 succeeds Claude Opus 4.8, and Anthropic says it comes close to the capability of Claude Fable 5. It targets everyday coding, knowledge work, and long-running agent workloads.

    Modalities
    text, image → text
    Context
    1M context
    At launch
    API, Consumer app, Cloud platforms
    • API model ID claude-opus-5; 1M-token context window and 128K max output tokens
    • New default model on Claude Max
    • Fast mode runs at around 2.5 times the default speed
    • Optional automatic fallbacks route safety-flagged API requests to another model instead of blocking them
    Source: Introducing Claude Opus 5
  11. MidjourneyImageProprietary

    Midjourney V8.2

    V8 update focused on aesthetics, image quality and personalization, producing bolder and more sophisticated images and reducing low-quality outliers, released as the default model.

    Modalities
    text, image → image
    At launch
    Consumer app
    • Released as the default version on July 24, 2026
    • Larger image pool for building personalization profiles
    • Followed on Aug 27, 2026 by an Edit Model that replaced Omni Reference
    Source: Version - Midjourney
  12. Black Forest LabsVideoProprietary

    FLUX 3 (Video)

    First FLUX multimodal foundation model, jointly trained on images, video and audio, announced in early access with video generation featuring synchronized audio, multilingual dialogue and keyframe control.

    Modalities
    text, image, video, audio → video, audio
    At launch
    API, Cloud platforms
    • Clips up to 20 seconds with native audio
    • Video generation became generally available via the BFL API on August 4, 2026
    • 2K and 4K output added to the video endpoint on September 10, 2026
    Source: FLUX 3 - Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence
  13. MicrosoftImageProprietary

    MAI-Image-2.5-Pro

    MAI-Image-2.5-Pro is Microsoft AI's highest-fidelity image model at launch, released in public preview for hero imagery, detailed editing and precise in-image text rendering, alongside the existing MAI-Image-2.5 and 2.5-Flash models.

    Modalities
    text, image → image
    At launch
    API
    • Released in public preview
    • Targets hero imagery, detailed editing and precise in-image text rendering
    • Sits above MAI-Image-2.5 and MAI-Image-2.5-Flash as the highest-fidelity option
    Source: Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash
  14. MicrosoftSpeechProprietary

    MAI-Voice-2-Flash

    MAI-Voice-2-Flash is a faster, lower-cost variant of Microsoft AI's MAI-Voice-2 text-to-speech model for high-volume voice experiences. First previewed at Microsoft Build in June 2026, it entered public preview in July 2026.

    Modalities
    text → audio
    At launch
    API
    • Released in public preview on July 23, 2026
    • Microsoft says it is 2x faster than MAI-Voice-2
    • Announced as coming soon on June 2, 2026
    Source: Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash
  15. GoogleLLMProprietary

    Gemini 3.5 Flash-Lite

    Gemini 3.5 Flash-Lite is the fastest and most cost-effective 3.5-class Gemini model, designed for high-throughput production traffic and as a low-latency subagent option for high-volume automation.

    Modalities
    text, image, audio, video → text
    Context
    1M context
    At launch
    API, Cloud platforms, Consumer app
    • Generally available in the Gemini API, Gemini Enterprise Agent Platform and the Gemini app, and rolling out in Google Search
    • Significant step up over prior Flash-Lite generations in agentic workflows
    • Input token limit of 1,048,576 per the Gemini API model page
    Source: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
  16. GoogleLLMProprietary

    Gemini 3.6 Flash

    Gemini 3.6 Flash is a Gemini workhorse model that improves coding, knowledge work and multimodal performance over 3.5 Flash while using fewer output tokens, fewer reasoning steps and fewer tool calls.

    Modalities
    text, image, audio, video → text
    Context
    1M context
    At launch
    API, Cloud platforms, Consumer app
    • Generally available in the Gemini API, Google Antigravity, Gemini Enterprise Agent Platform and the Gemini app
    • Reduces output verbosity in response to developer feedback on 3.5 Flash
    • Shipped with enhanced Frontier Safety safeguards for CBRN and cyber offense misuse
    Source: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
  17. Alibaba QwenImageProprietary

    Qwen-Image-3.0

    Qwen-Image-3.0 is the third-generation foundational image generation model in the Qwen-Image series. Qwen centers it on realism, with richer content, authentic details, and deeper knowledge.

    Modalities
    text, image → image
    At launch
    Consumer app
    • Supports up to 4.5K tokens of input for complex layouts such as newspapers and storyboards
    • Available to try on Qwen Chat
    • Successor to Qwen-Image-2.0
    Source: Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge
  18. ByteDance SeedSpeechProprietary

    Seed Audio 1.0

    Seed Audio 1.0 is an audio creation model that generates speech, sound effects, ambience, and other scene-level audio within one unified framework, coordinating character performance, timing, and background sound from a single prompt.

    Modalities
    text → audio
    At launch
    API
    • Models voice, sound effects, and ambient sound jointly instead of as separate clips
    • Keeps voice identity consistent across a broad expressive range
    • Available through BytePlus at launch
    Source: From Speech to Audio Creation | Introducing the Seed Audio 1.0 Audio Creation Model
  19. xAILLMProprietary

    Grok 4.5

    Grok 4.5 is SpaceXSI's model for coding, agentic tasks, and knowledge work, trained alongside Cursor on datasets spanning coding, science, engineering, and math across tens of thousands of NVIDIA GB300 GPUs.

    Modalities
    text, image → text
    Context
    500K context
    At launch
    API
    • API model ID: grok-4.5, with a 500k-token context window
    • Available at launch in Grok Build, Cursor on all plans, and the SpaceXSI console
    • Rolled out to grok.com, X, iOS, and Android on July 22, 2026
    • Configurable reasoning effort, default high
    Source: Introducing Grok 4.5 | SpaceXSI
  20. Moonshot AILLMOpen weights

    Kimi K3

    Kimi K3 is Moonshot AI's 2.8T-parameter model, which Moonshot describes as the first open 3T-class model. It is built on Kimi Delta Attention and Attention Residuals and targets long-horizon coding, knowledge work, and reasoning.

    Modalities
    text, image, video → text
    Context
    1M context
    Parameters
    MoE: 2.8T total, 104B active
    License
    Kimi K3 License
    At launch
    API, Consumer app, Open weights
    • Activates 16 of 896 experts per token
    • Natively understands text, images, and video
    • Launched on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full weights promised by July 27, 2026
    • Weights and code under the Kimi K3 License
    Source: Kimi K3: Open Frontier Intelligence
  21. MiniMaxMusicOpen weights

    MiniMax Music 3.0

    MiniMax Music 3.0 generates complete songs of up to five minutes from a creative description and optional lyrics, using a redesigned pipeline with an eight-layer RVQ tokenizer, a global-local Hybrid-LM, flow matching, and a Flow-VAE.

    Modalities
    text → audio
    Parameters
    8B global LLM plus 0.6B local LLM
    License
    MiniMax-Music3 Community License
    At launch
    API, Consumer app, Open weights
    • Listed in the MiniMax API release notes on July 16, 2026
    • Open weights published as MiniMax-Music3 with the August 13, 2026 announcement
    • Produces 32 kHz, 16-bit stereo audio
    • Music API closed to new users from August 20, 2026, with MiniMax pointing to MiniMax Audio and the open model
    Source: MiniMax Music 3.0: Next-Generation Open-Weights, Production-Ready & Versatile Music Model
  22. NVIDIAEmbeddingOpen weights

    Nemotron 3 Embed

    Text embedding models for retrieval and semantic similarity with multilingual and cross-lingual retrieval, built with bidirectional attention and average pooling.

    Modalities
    text → embedding
    Parameters
    1B and 8B
    License
    OpenMDW-1.1
    At launch
    Open weights
    • Released in 1B and 8B sizes
    • Intended as a foundation for RAG systems
    • Released under the OpenMDW 1.1 license
    Source: nvidia/Nemotron-3-Embed-8B-BF16 · Hugging Face
  23. TencentMultimodalOpen weights

    Hy-Embodied-VLM-1.0

    Hy-Embodied-VLM-1.0 is an open mixture-of-experts vision-language foundation model for embodied agents, built on the Hy3-A3B language backbone and the Hy-ViT2 vision encoder for latency-sensitive physical-world perception and reasoning.

    Modalities
    text, image → text
    Parameters
    MoE: ~30B total, ~3B active
    License
    Apache-2.0
    At launch
    Open weights
    • About 30B total parameters with roughly 3B activated per token
    • Activates about one-tenth the parameters of the previous A32B embodied system
    • Weights on Hugging Face
    Source: GitHub - Tencent-Hunyuan/HY-Embodied: HY-Embodied: Embodied Foundation Models for Real-World Agents
  24. Thinking Machines LabMultimodalOpen weights

    Inkling

    The first model Thinking Machines Lab released: an open-weights mixture-of-experts transformer that reasons natively over text, images and audio, with controllable thinking effort, built to be customized through Tinker.

    Modalities
    text, image, audio → text
    Context
    1M context
    Parameters
    MoE: 975B total, 41B active
    License
    Apache 2.0
    At launch
    API, Open weights
    • 975B total parameters, 41B active; pretrained on 45 trillion tokens of text, images, audio and video
    • Context window up to 1M tokens; Tinker offered 64K and 256K options at launch
    • Controllable thinking effort to trade performance against token use
    • Weights on Hugging Face (original and NVFP4) under Apache 2.0; served at launch by Together AI, Fireworks, Modal, Databricks and Baseten
    Source: Inkling: Our Open-Weights Model
  25. OpenAILLMProprietary

    GPT-5.6 (Sol, Terra, and Luna)

    GPT-5.6 is a model family with three durable tiers: Sol as the flagship, Terra balancing intelligence and cost, and Luna for efficient high-volume work, adding a max reasoning effort and an ultra multi-agent mode.

    Modalities
    text, image → text
    Context
    1.1M context
    At launch
    Consumer app, API, Enterprise
    • API model IDs gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna; the gpt-5.6 alias routes to Sol
    • General availability followed a limited June 26, 2026 preview for a small group of trusted partners
    • In ChatGPT, Plus and higher plans use Sol, with GPT-5.6 Sol Pro for Pro and Enterprise users
    • All three tiers list a 1,050,000-token context window on their API model pages
    Source: GPT-5.6: Frontier intelligence that scales with your ambition | OpenAI
  26. MetaReasoningProprietary

    Muse Spark 1.1

    Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, upgrading Muse Spark in tool and computer use, coding, multi-agent orchestration and context management.

    Modalities
    text, image, video, audio → text
    Context
    1M context
    At launch
    API, Consumer app
    • Available to developers through the Meta Model API public preview
    • Manages a context window of 1 million tokens
    • Available in Thinking mode in the Meta AI app and on meta.ai
    Source: Introducing Muse Spark 1.1
  27. OpenAISpeechProprietary

    GPT-Live-1 and GPT-Live-1 mini

    GPT-Live is a new generation of full-duplex voice models that can listen and speak at the same time, delegating search and deeper reasoning to a frontier model in the background, and it now powers ChatGPT Voice.

    Modalities
    audio, text → audio, text
    At launch
    Consumer app
    • Rolled out to ChatGPT users globally: GPT-Live-1 for paid users and GPT-Live-1 mini for Free users
    • At launch it delegated harder requests to GPT-5.5 in the background
    • GPT-Live-1 became generally available in the API as gpt-live-1 on September 10, 2026
    • Not available in ChatGPT Business, Enterprise, or Edu workspaces at launch
    Source: Introducing GPT-Live | OpenAI
  28. ByteDance SeedImageProprietary

    Seedream 5.0 Pro

    Seedream 5.0 Pro is a multimodal image creation model aimed at professional design work, adding complex information visualization, interactive region-based editing, layer separation, and native multilingual text rendering to the Seedream line.

    Modalities
    text, image → image
    At launch
    API, Consumer app
    • Supports point, lasso, box, and sketch selections for precise local edits
    • Can split a poster into more than 10 editable layers
    • Generates text natively in over ten languages beyond Chinese and English
    • Launched in Volcano Ark's experience center, rolling out to Doubao and Jimeng
    Source: Beyond Generation, It Understands Design | Introducing Seedream 5.0 Pro
  29. CohereSpeechOpen weights

    Cohere Transcribe Arabic

    Open-source speech-to-text model that fine-tunes Cohere Transcribe on Arabic speech data to transcribe all major Arabic dialects as well as English spoken with an Arabic accent.

    Modalities
    audio → text
    Parameters
    2B
    License
    Apache-2.0
    At launch
    API, Open weights
    • Conformer-based encoder-decoder architecture
    • Model name cohere-transcribe-arabic-07-2026
    • Open weights on Hugging Face
    Source: Meet Cohere Transcribe Arabic
  30. TencentMultimodalOpen weights

    HunyuanOCR-1.5

    HunyuanOCR-1.5 upgrades Tencent's lightweight end-to-end OCR vision-language model without changing its backbone, adding DFlash speculative decoding, PC-side llama.cpp deployment, and an upgraded data and training recipe.

    Modalities
    text, image → text
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Covers document parsing, text spotting, information extraction, and text-image translation
    • Ships a DFlash block-diffusion draft model for faster decoding
    • HunyuanOCR-1.0 weights remain archived in the same repository
    Source: GitHub - Tencent-Hunyuan/HunyuanOCR: HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better
  31. MetaImageProprietary

    Muse Image

    Muse Image is Meta Superintelligence Labs' first image generation model, an agentic generator and editor that uses search and coding tools, refines its own outputs and composes from multiple references.

    Modalities
    text, image → image
    At launch
    Consumer app, API
    • Available in the Meta AI app, on meta.ai, and in Instagram Stories in the US at launch
    • Supports multi-turn precise editing and multi-reference composition
    • Outputs carry Content Seal, an invisible watermark
    • Announced with Muse Video, which was only previewed
    Source: Introducing Muse Image and Muse Video
  32. OpenAISpeechProprietary

    gpt-realtime-2.1 and gpt-realtime-2.1-mini

    GPT-Realtime-2.1 updates GPT-Realtime-2 with improved alphanumeric recognition, silence and noise handling, and interruption behavior, and was released alongside GPT-Realtime-2.1 mini, a faster, lower-cost distilled reasoning model for realtime voice.

    Modalities
    text, audio, image → text, audio
    Context
    125K context
    At launch
    API
    • API model IDs gpt-realtime-2.1 and gpt-realtime-2.1-mini in the Realtime API
    • Both list a 128,000-token context window and 32,000 max output tokens
    • The mini model supports audio and text input over WebRTC, WebSocket, or SIP connections
    Source: Changelog | OpenAI API
  33. TencentLLMOpen weights

    Hy3

    Hy3 is the full release following Hy3 preview, with scaled-up post-training and reinforcement learning that improve reasoning, agentic, and long-context tasks, plus fixes for tool-call stability and hallucination based on feedback from over 50 products.

    Modalities
    text → text
    Context
    250K context
    Parameters
    MoE: 295B total, 21B active
    License
    Apache-2.0
    At launch
    Open weights, API
    • 295B total and 21B active parameters, plus 3.8B MTP layer parameters
    • 256K context length
    • Weights on GitHub, Hugging Face, ModelScope, and AtomGit
    • API on Tencent Cloud and OpenRouter
    Source: Introducing Hy3
  34. Mistral AICodingOpen weights

    Leanstral 1.5

    Updated Lean 4 proof engineering model trained through mid-training, supervised fine-tuning and reinforcement learning with CISPO, aimed at automated theorem proving and code verification.

    Modalities
    text → text
    Context
    250K context
    Parameters
    MoE: 119B total, 6B active
    License
    Apache-2.0
    At launch
    Open weights, API
    • API access began June 30, 2026 as labs-leanstral-1-5
    • Trained in multi-turn proving and code agent environments
    • Retired from the API September 30, 2026
    Source: Leanstral 1.5: Proof Abundance for All

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 July 2026." The AI Engineering, updated Jul 31, 2026. https://theaiengineering.com/releases/2026/07

HTML link

<a href="https://theaiengineering.com/releases/2026/07">AI models released in July 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

Archive