New AI Models

Seed Audio 1.0

Seed Audio 1.0 is ByteDance's audio creation model for full-scene audio generation. It jointly models voice, sound effects, and ambience in a unified framework for end-to-end film-grade audio production, with prompt-level timing control at 100 ms intervals. The model supports zero-shot voice generation from text or reference samples, produces up to two minutes of audio in a single pass, and transfers the same voice naturally across more than 20 languages. Official evaluations show an audio Availability Rate above 90% in most scenarios such as film, short dramas, and podcast dialogue.

Qwen-Image-3.0-Pro

Qwen-Image-3.0-Pro is an image generation model available on Alibaba Cloud Model Studio, positioned as a practical productivity tool rather than a purely aesthetic one. It accepts up to 4.5k tokens of input, generating complex layouts such as newspapers, storyboards, menus, and exam papers in a single pass. It renders text as small as 10px with precision and reproduces fine details like micro-expressions and hair strands with near-photographic realism, while natively rendering text in 12 languages and simulating web, game, and livestream interfaces.

Qwen3.7-Flash

Qwen3.7-Flash is the fast, cost-effective tier of Alibaba's Qwen3.7 family, a natively vision-language model that accepts image, text, and video input. Compared with Qwen3.6-Flash, it delivers stronger multimodal understanding and object recognition, improved real-world perception and spatial intelligence, and more stable end-to-end execution in agentic scenarios such as Search Agent and CI Agent. Multimodal coding and the vibe coding experience are also refined. It offers a 1M context window, with input from RMB 0.2 per million tokens in the Beijing region.

Qwen3.7-Flash-2026-07-15

Qwen3.7-Flash-2026-07-15 is the 2026-07-15 snapshot of Qwen3.7 Flash, the fast and cost-efficient tier of Alibaba Cloud's natively multimodal Qwen3.7 vision-language series on Model Studio, suited to production deployments that need a pinned version. Compared with 3.6-Flash it delivers stronger multimodal understanding and agent execution, better object recognition and spatial intelligence, and accepts image, text, and video input with a 1M-token context. Typical workloads include Search Agent and CI Agent scenarios as well as multimodal coding and vibe coding.

Gemini 3.5 Flash Cyber

Gemini 3.5 Flash Cyber is Google's lightweight cybersecurity model, fine-tuned from Gemini 3.5 Flash to find, validate, and patch vulnerabilities more effectively than the mainline Flash models. Its speed and affordability suit scanning large codebases across many code paths. It proved competitive with significantly larger models on the CyberGym benchmark and surpassed mainline 3.5 Flash and 3.6 Flash on Google's Big Sleep evaluation. It is offered to governments and trusted partners via CodeMender through a limited-access pilot.

MiniMax Music-3.0

MiniMax Music-3.0, released July 16, 2026, is MiniMax's latest music generation model. An upgraded semantic model better captures creative intent and reduces AI-generated drift, while a new vocal engine delivers natural singing with control over melody, pronunciation, breathing, and layered harmonies. It supports specific instruments and real playing techniques such as slides and legato for commercial-recording-grade sound, and can restyle existing songs as covers. Accessible via the Music Generation API, it suits theme songs for videos, games, and apps.

Mistral OCR 4.1

Mistral OCR 4.1 is the latest version of Mistral AI's OCR service powering its Document AI stack, released on July 16, 2026. It succeeds OCR 4, taking over the mistral-ocr-latest and mistral-ocr-4 aliases, and adds native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores. Served through the Mistral OCR API with structured annotations and batch processing support, it is priced at $4 per 1,000 pages ($5 per 1,000 annotated pages), fitting document digitization, parsing, and structured extraction workflows.

Kimi K3

Kimi K3 is Moonshot AI's open-weight, native multimodal agentic model and its most capable release to date. Built on a 2.8T-parameter MoE architecture with 104B activated parameters and a 1M-token context window, it handles text, image, and video in one model. Moonshot positions it as the world's first open 3T-class model, competitive with GPT-5.6 Sol and Claude Opus 4.8 on benchmarks such as GPQA Diamond, SWE-Marathon, and BrowseComp. It ships under the Kimi K3 License for long-horizon coding and agentic knowledge work.

Doubao-Seed-Evolving

Doubao-Seed-Evolving is ByteDance Seed's self-evolving model line built for Agent and Coding scenarios. Instead of frozen version snapshots, it ships through one permanent Model ID, doubao-seed-evolving: updates roll out on a weekly cadence and take effect automatically, with no endpoint migration or code changes. The current version offers a 1M-token context window, deep thinking, function calling, and image and video understanding. It is served on Volcano Engine's Ark platform via OpenAI- and Anthropic-compatible APIs, working in tools like Claude Code.

Qwen-Audio-3.0-TTS-Flash

Qwen-Audio-3.0-TTS-Flash is a high-performance text-to-speech model from Alibaba's Qwen team, optimized for real-time interaction. It supports more low-resource languages and Chinese dialects than its predecessor, with free-style instruction following and fine-grained tags for controlling emotion, tone, character, speech rate, and volume. The model is also more robust in noisy and reverberant acoustic conditions. The Flash tier prioritizes real-time synthesis, making it a fit for voice assistants, real-time conversation, and intelligent customer service.

Qwen-Audio-3.0-TTS-Plus

Qwen-Audio-3.0-TTS-Plus is a high-performance text-to-speech model from Alibaba's Qwen team built for high-quality speech generation. It supports more low-resource languages and Chinese dialects, with free-style instructions and fine-grained tags for controlling emotion, tone, speaking rate, and speaking style. It is also more robust under noise and reverberation, with further improved audio quality, clarity, and expressiveness. The Plus tier prioritizes output quality and detail, making it a fit for content creation, audiobooks, film dubbing, and brand voice design.

Muse Spark 1.1

Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs, built for agentic tasks with major gains in tool and computer use, coding, and multimodal understanding. It actively manages a 1 million token context window, orchestrates multi-agent systems to cut end-to-end latency, and posts substantial coding gains over its predecessor on Meta Internal Coding Bench. Developers can access it through the Meta Model API public preview, or use it in Thinking mode in the Meta AI app and on meta.ai.

GPT-5.6

OpenAI's GPT-5.6 family of frontier models ships in three capability tiers: Sol, the new flagship; Terra, balanced for everyday work; and Luna, a fast, cost-efficient option, available across ChatGPT, Codex, and the OpenAI API. Sol delivers state-of-the-art results in coding, knowledge work, cybersecurity, and science, outperforming GPT-5.5 and Claude Fable 5 with fewer tokens and lower estimated cost. A new ultra setting coordinates multiple agents in parallel to finish demanding tasks faster.

Seedream 5.0 Pro

Seedream 5.0 Pro is a multimodal image generation model from ByteDance's Seed team, built for advanced reasoning and professional production. It handles high-density infographics with rich, accurate text rendering and supports interactive editing that follows spatial annotations and sketches, including layer separation. With authentic lighting and skin textures plus native prompting in a dozen widely used languages, it fits education graphics, poster design, and multilingual content workflows.

Robostral Navigate

Robostral Navigate is Mistral AI's first embodied navigation model: an 8B model that guides a robot through unfamiliar environments from plain-language instructions and a single RGB camera, with no LiDAR or depth sensors. Built in-house and trained entirely in simulation on roughly 2.4 million trajectories, it reaches 76.6% success on R2R-CE validation unseen, beating the best single-camera approach by 9.7 points. It generalizes across wheeled, legged, and flying robots, targeting manufacturing, delivery, logistics, and hospitality.

GPT-Live

GPT-Live is OpenAI's new real-time voice model built on a full-duplex architecture, so it listens and speaks at the same time and handles natural turn-taking, interruptions, and live translation. It is OpenAI's smartest voice model to date, delegating search and deeper reasoning to a frontier model in the background (GPT-5.5 at launch) while the conversation keeps flowing. In human evaluations it was clearly preferred over Advanced Voice Mode. GPT-Live-1 and GPT-Live-1 mini are rolling out to ChatGPT users worldwide on iOS, Android, and the web, with API access planned.

Muse Image

Muse Image is the first image generation model from Meta Superintelligence Labs, launched on July 7, 2026 as an agentic image generator that follows instructions faithfully, edits with precision, and composes from multiple reference images. It invokes search and coding tools, self-refines its outputs, and improves with test-time compute, ranking No. 2 on Arena for text-to-image and image editing at launch. It is available in the Meta AI app, on meta.ai, Instagram Stories in the US, and WhatsApp in limited countries, with a Content Seal watermark on outputs.

Muse Video

Muse Video is one of the first media generation models from Meta Superintelligence Labs, previewed on July 7, 2026 alongside Muse Image and built on the same pretraining base. It generates video with native audio support and competitive prompt adherence, visual fidelity, and temporal consistency, ranking No. 3 on Arena's text-to-video Elo leaderboard at publication. Currently in early preview ahead of a wider rollout to creators and Meta AI, with Content Seal watermarking planned to extend to video.

hy3

Hy3 is a reasoning and agent model from Tencent Hunyuan, built on a MoE architecture with 295B total parameters, 21B activated per token, and a 256K context window. Tencent positions it as rivaling flagship models with 2-5x its parameters, with particular strength in software development, office productivity, and financial modeling. The model is open-sourced under Apache 2.0 and available through Tencent Cloud API.

Vidu S1

Vidu S1, released by Shengshu Technology (Vidu) on July 3, 2026, is a real-time interactive video generation model. It streams video at 540P and 25 FPS (up to 42 FPS) and is billed as the first generative video model to support unlimited interactive session length. Users interact through voice commands and video calls, controlling digital human behavior in real time, with customizable avatars (real person, anime, pet) and voice cloning. It is available on the web and via API through Vidu's MaaS platform.