New AI Models

Qwen-Audio-3.1-Realtime-Plus

Qwen-Audio-3.1-Realtime-Plus is an end-to-end realtime speech-to-speech model from Alibaba's Qwen team for low-latency duplex voice interaction, with streaming audio input and streaming speech or text output. It follows the 3.0 Plus access protocol over WebSocket, AOQ and WebRTC, keeps the existing voices and adds 8 new system voices, with a 262,144-token context and support for Function Calling, web search and voice cloning. Its smart_turn mode judges turn boundaries from acoustic and semantic cues, so filler sounds like uh and um do not interrupt the dialogue. It supports 11 languages and about 21 Chinese dialects, fitting voice assistants, customer service bots and AI companions.

Qwen-Image-2.1-PE-I2I

Qwen-Image-2.1-PE-I2I is an image editing prompt rewriting model for Qwen-Image-2.1, fine-tuned from Qwen3.5-VL 9B and positioned as the image-to-image counterpart of the text-to-image oriented PE-T2I. It takes a vague editing instruction together with one or more input images and rewrites them into a precise, actionable prompt, outputting JSON that carries a rewritten_prompt plus either a target aspect ratio (wh_ratio) or a flag to keep the input ratio (ratio_follow). The rewritten prompt plugs directly into the QwenImage21Pipeline in Diffusers, making the model a helper for building reliable automated image editing flows. It is released under the Qwen Research License.

Qwen-Image-2.1-PE-T2I

Qwen-Image-2.1-PE-T2I is a prompt rewriting companion model for the Qwen-Image-2.1 text-to-image system, fine-tuned from Qwen3.5-VL 9B. It turns a brief image request in any language into a detailed English prompt plus a recommended aspect ratio, returned as JSON with rewritten_prompt and wh_ratio fields that feed directly into the QwenImage21Pipeline in diffusers. Generation runs with thinking mode enabled, and the model is released under the Qwen Research License.

Grok Voice Transcribe 2.0

Grok Voice Transcribe 2.0 is a speech-to-text model built on the audio foundation model behind Grok Voice, trained on live, noisy, multilingual audio from real-world environments such as flaky phone lines, competing speakers, and local accents. It is claimed to be twice as accurate as Grok Voice Transcribe 1.0 at the same price, ranking first among 32 streaming models on the Artificial Analysis leaderboard, with word error rate on short multilingual commands dropping from 20.6% to 6.8%. Features include batch and streaming transcription, word-level timestamps, speaker diarization, up to 8 channels, key term biasing, and smart turn detection, priced at $0.10 per hour for batch and $0.20 for streaming. It targets customer-support calls, voice agents, and large-scale video transcription.

HappyOyster-1.0-Acting

HappyOyster-1.0-Acting is the character acting mode of Alibaba Cloud Model Studio's HappyOyster real-time interactive world model series. It takes a natural language prompt defining a persona plus a first frame image (default 9:16 portrait, 16:9 supported) and renders the character and scene as a live video stream. Text instructions sent in real time drive the character's performance and interaction, with pause and resume support, making it suited to video-call style AI companion apps and other portrait-oriented character experiences.

HappyOyster-1.0-Adventure

HappyOyster-1.0-Adventure is a real-time interactive, explorable open-world model from Alibaba's Qwen team. It generates world scenes from text or image input, and supports real-time interaction, movement, and camera control so users can roam each generated world without boundaries. Output is delivered as video, which fits interactive world exploration experiences and world-building content creation.

HappyOyster-1.0-Directing

HappyOyster-1.0-Directing is the directing variant of Alibaba Cloud Model Studio's HappyOyster world model line, alongside the Adventure and Acting variants. It is a real-time interactive, immersively performable open-world model that generates world scenes from multimodal text and image input. At any moment, text prompts can drive the world to unfold and evolve in real time, putting you in the director's seat of a living scene. It suits interactive storytelling, open-world creation, and immersive playable experience scenarios.

Qwen3.8-Omni-Flash

Qwen3.8-Omni-Flash is a native omni-modal model on Alibaba Cloud Model Studio that accepts text, image, audio, and video input and produces text output. It supports up to 1M context tokens, takes two-channel stereo and four-channel FOA spatial audio via a multichannel switch, and provides adjustable thinking and non-thinking modes alongside Function Calling and web search. It fits audio and video understanding, meeting summaries, content analysis, and agentic applications, while live interaction is handled by its Realtime sibling.

Qwen-MT-Uni

Qwen-MT-Uni is a unified multimodal translation model from the Qwen team, served via Alibaba Cloud Model Studio, that translates text, documents (PDF, Word, PPT, Excel, HTML, Markdown, TXT), images, and audio in a single API call. The server automatically detects the input modality and runs a pipeline of format recognition, intelligent routing, content extraction, cross-modal translation, and reconstruction of the original layout, with voice cloning for audio output. It supports both synchronous and asynchronous invocation, accepts files up to 100 MB (up to 200 pages per document, audio from 3 seconds to 60 minutes), and returns translated text or a file download link, making it suitable for large-scale cross-modal translation workloads.

Seed 2.1 Lite

Doubao Seed 2.1 Lite is the lightweight tier of ByteDance's Seed 2.1 family, a new-generation omni-modal model with deep thinking capability. It natively unifies understanding of text, vision, and audio, and supports a 1M token context window for long-context tasks. ByteDance reports strong overall multimodal understanding, with notable gains in Coding and Agent task completion. It fits cost-sensitive applications that need multimodal understanding, long context, and coding or agentic workflows.

Seedream 5.0 Flash

Doubao Seedream 5.0 Flash is ByteDance's fast, low-cost image generation model on Volcano Ark, matching Seedream 5.0 Pro in capability while generating images faster and at a lower price. It supports text-to-image, single-image and multi-reference image-to-image, plus interactive editing that targets edits via coordinates, box selection, or arrows for local element replacement and regional generation. It can split an image into one base layer and up to 16 independent transparent-channel layers, and natively renders text in 14 languages, making it a fit for latency- and cost-sensitive workflows.

Gemini 3.8 Live

Gemini 3.8 Live (gemini-3.8-live) is Google's audio-to-audio live dialogue model served through the Live API, positioned by the company as the default choice for most low-latency voice agent experiences and real-time conversations, with no added inference delay. Google describes it as built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding: it processes near real-time visual input, switches automatically among 97 supported languages mid-conversation, and runs tool and API calls in the background while the conversation continues. Unlike its Extended Thinking sibling, which targets multi-step reasoning on high-complexity tasks, this version focuses on responsive dialogue, and it is available in the Gemini API, Google AI Studio, the Gemini app and Search Live, with generated audio watermarked via SynthID.

Gemini 3.8 Live Extended Thinking

Gemini 3.8 Live Extended Thinking is Google's high-complexity variant of the Gemini 3.8 Live dialogue family, adding an extended thinking mode that reasons and speaks simultaneously. It runs tool and API calls in the background while the conversation continues, narrates multi-step progress, and grounds replies in live visual input across 97 languages. Google reports it tops Artificial Analysis' Speech to Speech Quality Index at 82.6 and leads agentic voice benchmarks such as tau-Voice at 68.6%. It is offered via the Gemini API, AI Studio, and the Gemini app, with all generated audio watermarked by SynthID.

Vidu S2-Avatar

Vidu S2-Avatar is a real-time interactive digital-character video generation model from Shengshu AI, part of the Vidu S2 family. It generates 720p avatar video at 25 to 42 FPS, follows a wider range of instructions including large body motions such as dancing, and accepts new reference images at any moment during a stream. It is suitable for live interactive avatars, digital human companions, and real-time character driven applications, and is available to try on vidu.com.

Vidu S2-Editing

Vidu S2-Editing is a real-time video editing model that Vidu offers to enterprise customers: it continuously receives a published video stream and outputs an edited stream as frames arrive, integrated via an HTTP, WebSocket and WebRTC (AliRTC) API. It supports four editing capabilities driven by an editing scenario with an optional reference image: style transfer, virtual try-on, subject replacement, and background replacement, each preserving the motion, expressions, and camera movement of the source stream. The reference image or scenario can be switched mid-stream without interrupting the video, making it suitable for stylized live streaming, virtual try-on showcases, and interactive character or background swap broadcasts.

Music v2.5

Music v2.5 (model_id music_v2_5) is ElevenLabs' most advanced music generation model, producing studio-grade songs from text prompts, composition plans, or previously generated tracks. The official docs call out improved quality and prompt adherence over music_v2, with richer melodies, deeper arrangements, and more layered instruments, while keeping the same workflows: composition plans, audio reference, and inpainting. Vocals are available in languages including English, Spanish, German, and Japanese, with typical uses spanning game soundtracks, podcast backgrounds, and marketing. The model works across the music API endpoints and SDKs, and composition plan generation chunks now allow up to 6,132 characters.

Qwen-Image-2.1

Qwen-Image-2.1 is an open-source unified text-to-image generation and image editing model from the Qwen family, with a 7B visual generation component built on 32 Single-Stream DiT layers that balances quality and cost through mixed-granularity attention and prefix KV cache reuse. It natively generates and edits RGBA transparent images, extracts subjects from photos, accepts up to 10 reference images with identity preservation for people and products, and supports local edits specified by circles, painted annotations, or separate masks at native 2K resolution. Improved typography, portrait lighting, and fine details make it suitable for design, e-commerce, and visual storytelling, with Day-0 support in diffusers, ComfyUI, vLLM-Omni, and SGLang.

Scribe v2 Medical

Scribe v2 Medical is a batch speech recognition model specialized for medical and clinical audio, now generally available via model_id scribe_v2_medical on the same Speech to Text API as Scribe v2. As a fine-tune of Scribe v2, it improves recognition of drug names, anatomy, pathology and clinical dictation, with 15% fewer errors than Scribe v2 on isolated medical terms and no regression on everyday speech. It keeps the same request shape as Scribe v2: keyterm prompting up to 1000 terms, entity detection including PHI categories, speaker diarization, no-verbatim mode and word-level timestamps across 90+ languages, billed at the same rate. Typical uses include ambient clinical documentation, medical dictation, patient intake calls and HIPAA compliance workflows, with a BAA available through ElevenLabs sales.

DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is the smallest model in DeepSeek's new CED (Causal Encoder-Decoder) architecture family: a 552B-parameter MoE with native visual understanding that activates 8B parameters for input and 16B for output. It supports a 1M-token context, and its FP4 KV cache needs about 1/4 the HBM and 1/8 the SSD storage of the previous generation. Official evaluations report GPQA Diamond 90.9, HLE 36.8, Codeforces 3471, Terminal-Bench 2.1 90.6 and DeepSWE v1.1 74.2, beating its flagship DeepSeek-V4-Pro on several of these benchmarks. Served as deepseek-flash with lower launch pricing, it replaces the retired V4-Flash and V4-Flash-Vision-Exp (old names route temporarily), and its MIT-licensed weights fit high-throughput multimodal, coding and agentic workloads.

GPT-Live-1

GPT-Live-1 is OpenAI's full-duplex real-time voice model, now generally available in the API after first appearing in ChatGPT. A single model handles listening and speaking at the same time, reasoning over incoming and outgoing audio to avoid the latency and brittle handoffs of chained STT-LLM-TTS pipelines, and it can delegate deeper reasoning and tool calls to a backend model such as GPT-6 Astra or a third-party option. OpenAI reports a 30 percentage point Full Duplex Bench improvement over GPT-Realtime-2.1 and a #1 rank on Tau3 when paired with GPT-6 Astra at medium reasoning effort, with telephony support, native ASR transcripts, and improved interruption handling for phone and in-app voice agents. Pricing is $0.05 per minute for the front-end voice layer.