New AI Models

kling-v2-1

kling-v2-1 is a video generation model on the Kling AI open platform by Kuaishou, part of the Kling 2.1 generation alongside the higher tier kling-v2-1-master. It is documented for image-to-video generation with two modes: standard mode for cost effective 720P output and professional mode for higher quality 1080P output, producing 5 or 10 second clips, with start and end frame control available in professional mode. Positioned as an iteration on Kling 2.0, it improves visual quality, motion plausibility and prompt adherence. The 2.1 generation also covers text-to-video through the master variant. It suits short form marketing clips, social media content and storyboard previews, and official docs now list it as an earlier generation model scheduled for retirement as newer Kling releases take over.

kling-v2-new

kling-v2-new is the API model identifier for the Kling 2.0 generation from Kuaishou's Kling AI. The Kling 2.0 video model was officially released globally in April 2025 and supports text-to-video and image-to-video, with stated upgrades in dynamic quality, semantic response and visual aesthetics: stronger prompt adherence, a more cinematic look, and control over creative dimensions such as scene, action, expression and camera movement. In Kuaishou's internal multi-metric evaluation, its image-to-video results recorded win-loss ratios of 182% against Google Veo 2 and 178% against Runway Gen-4. The generation also introduced the Multi-modal visual language (MVL) interaction concept, multimodal video editing for adding, removing or replacing elements, over 60 stylized effect transfers, and API access for enterprise customers.

Gemini 2.5 Pro

Gemini 2.5 Pro is Google's thinking flagship multimodal model, built to reason through its thoughts before responding. The May 2025 I/O edition Preview greatly improves coding, especially interactive web apps, and extends to code transformation, code editing and complex agentic workflows, leading the WebDev Arena Leaderboard with a gain of +147 Elo points over its predecessor. It keeps native multimodality and a 1 million token context window (2 million to follow), scores 84.8% on the VideoMME benchmark for state-of-the-art video understanding, and debuted at #1 on LMArena. Developers can build with it via the Gemini API in Google AI Studio and Vertex AI, and it powers the Gemini app for complex tasks.

V6

V6 is an image generation model version from Midjourney. Per the official documentation, it offers enhanced prompt accuracy for longer inputs, improved coherence and knowledge, and advanced image prompting and remixing capabilities. It works with style reference, image prompts, variations, pan, zoom out, and the Subtle and Creative upscalers, and supports aspect ratios up to 14:1, which suits illustration, concept art, and brand visuals that depend on precise prompt adherence. Its successor V6.1 later delivered more coherent details and faster generation.

v4.5

Suno v4.5 is an AI music generation model that creates songs up to 8 minutes long while keeping quality and coherence from start to finish. It offers richer vocals with wider range and emotional depth, a much larger genre selection with better genre mashups, and more complex sound details such as tone shifts and instrument layering. It also understands prompts more accurately, adds a prompt enhancement helper, upgrades Covers and Personas so they can be combined, and generates faster with less audio degradation and shimmer.

v4.5-all

v4.5-all is Suno's music generation model opened up to every user, including the free tier, bringing the full v4.5 capability set to all plans. It supports songs up to 8 minutes, vocals with wider range and more emotional depth, more accurate genres and better genre mashups, smarter prompt interpretation with a prompt enhancement helper, upgraded Covers and Personas, faster generation, and fuller mixes with less audio degradation. It fits songwriting, genre experiments, and producing complete tracks.

Phi 4 Mini Reasoning

Phi-4-mini-reasoning is a 3.8B parameter open-source reasoning model from Microsoft, built on Phi-4-mini and fine-tuned on more than one million synthetic math problems distilled from DeepSeek-R1 with verified chain-of-thought solutions. Despite its small size, it delivers strong multi-step mathematical reasoning: the official model card reports 94.6 on MATH-500, ahead of OpenAI o1-mini at 90.0, plus 57.5 on AIME and 52.0 on GPQA Diamond, outperforming 7B and 8B reasoning distills of similar or larger scale. Released under the MIT license with a 128K token context, it targets memory and compute constrained environments such as edge and mobile devices, educational apps, and embedded tutoring that need accurate step-by-step problem solving at low latency.

Phi 4 Reasoning

Phi-4-reasoning is a 14B parameter open-weight reasoning model from Microsoft, released on April 30, 2025. It was built from Phi-4 via supervised fine-tuning on curated chain-of-thought data for complex reasoning, and it answers with a reasoning block followed by a summarized solution. It handles a 32k token context, ships under the MIT license, and is available on Azure AI Foundry and Hugging Face. On official benchmarks it scores 62.9 on AIME 2025, 76.6 on OmniMath, and 65.8 on GPQA-Diamond, outperforming the far larger DeepSeek-R1-Distill-70B and approaching the full DeepSeek-R1 model. It fits memory or compute constrained deployments, latency bound scenarios, math and science reasoning workloads, and agentic applications that need a reasoning backbone.

Phi 4 Reasoning Plus

Phi-4-reasoning-plus is a 14B open-weight reasoning model from Microsoft, released on April 30, 2025 under the MIT license. It is built on Phi-4 with supervised fine-tuning on curated chain-of-thought demonstrations, then further trained with reinforcement learning to use more inference-time compute. Compared with Phi-4-reasoning it delivers higher accuracy at the cost of generating about 50% more tokens and higher latency. Microsoft reports AIME 2025 at 78.0 (versus 62.9 for Phi-4-reasoning), AIME 2024 at 81.3, and OmniMath at 81.9, with both variants beating DeepSeek-R1-Distill-Llama-70B on most benchmarks and outscoring the full DeepSeek-R1 on AIME 2025. It targets math and logic heavy reasoning in memory or compute constrained environments, with 32k context (tested up to 64k).

Qwen-Plus-2025-04-28

Qwen-Plus-2025-04-28 is a dated snapshot of Qwen-Plus, the balanced commercial text model on Alibaba Cloud Model Studio, released on April 29, 2025. Built on the Qwen3 series, it integrates thinking and non-thinking modes in one model and lets you switch between them during a conversation. Per the official documentation, its reasoning notably surpasses QwQ and its general capabilities notably surpass Qwen2.5-Plus, reaching SOTA level among models of its scale at release. It offers a 131,072-token context window, function calling, structured output, and prefix continuation, fitting general purpose workloads such as complex reasoning, tool use, and production text generation.

Qwen3-1.7B

Qwen3-1.7B is a compact dense model from the Qwen3 open-weight series released on April 29, 2025 under Apache 2.0. It has 1.7B parameters (1.4B non-embedding) across 28 layers with GQA (16 query heads, 8 KV heads) and a native context of 32,768 tokens. A single model switches between thinking mode for step-by-step reasoning on hard problems and non-thinking mode for quick replies, controlled via the chat template or /think and /no_think soft switches. The series was pretrained on about 36 trillion tokens, and the model supports 119 languages and dialects with strong multilingual instruction following, translation, coding, math, and agentic tool use. Per the official blog, its base version performs on par with the larger Qwen2.5-3B-Base, making it well suited to local and low-resource deployment through Ollama, llama.cpp, LMStudio, and MLX.

gpt-image-1

gpt-image-1 is OpenAI's natively multimodal image generation model, the API version of the image generation feature in ChatGPT. It accepts text and image inputs and produces image outputs, supporting text-to-image generation, image editing, accurate text rendering, and the use of reference images. It creates images across diverse styles, follows custom guidelines faithfully, and leverages world knowledge to interpret prompts. Companies including Adobe, Canva, GoDaddy, HubSpot, Instacart, and invideo use it for design, e-commerce, marketing, and education scenarios. It shares the safety guardrails of ChatGPT image generation, adds C2PA metadata to outputs, and offers a moderation parameter for developers; usage is priced per token, roughly $0.02 to $0.19 per square image depending on quality.

Protenix

Protenix is an open-source biomolecular structure prediction system from ByteDance, built as a trainable AlphaFold3 reproduction that predicts 3D structures of protein complexes with nucleic acids, ligands, and other molecules. Code and model parameters are released under Apache 2.0, together with the full training and MSA pipelines. According to the official benchmark, Protenix-v1 is the first fully open-source model to outperform AlphaFold3 across diverse benchmark sets under the same data cutoff, model scale, and inference budget, while Protenix-v2 adds 9 to 13 percentage points of absolute success rate over v1 at the DockQ > 0.23 threshold on antibody-antigen collections. It fits drug discovery and life science workflows such as complex structure modeling, protein-ligand interaction analysis, and antibody engineering, and offers a web server plus lightweight Protenix-Mini variants for lower-cost inference.

Seed GR-3

Seed GR-3 is ByteDance Seed's large-scale Vision-Language-Action (VLA) model for robot manipulation. Trained on a multimodal foundation, it generalizes to unseen objects, environments, and instructions involving abstract concepts, and supports efficient fine-tuning with a small number of human demonstrations collected via VR devices, allowing quick and low-cost adaptation to new tasks. GR-3 performs robustly on long-horizon and dexterous tasks that require bimanual manipulation and mobile base movement, such as table bussing and clothes hanging, and is deployed on ByteMini, the team's bimanual mobile robot. It targets real-world scenarios like household chores, object rearrangement, and deformable-object handling; the technical report is available on arXiv (2507.15493).

Seed LiveInterpret 2.0

Seed LiveInterpret 2.0 is ByteDance Seed's end-to-end speech-to-speech simultaneous interpretation model for bidirectional Chinese and English. Built on a full-duplex speech understanding and generation framework, it offers zero-shot voice cloning that keeps each speaker's timbre, and stays stable in multi-speaker conversations, disfluent speech, and long audio. In official human evaluations it scored 74.8 out of 100 for speech-to-text translation quality (versus 47.3 for the second-ranked baseline) and 66.3 out of 100 for speech-to-speech output, a level close to professional human interpreters, with average first-output latency of 2.21 seconds for speech-to-text and 2.53 seconds for speech-to-speech. It targets live scenarios such as remote meetings and news broadcasting, and is available via Volcano Engine with a technical report on arXiv.

Seed Realtime Voice

Seed Realtime Voice (officially introduced by ByteDance Seed as SeedRealtime) is a native audio-visual full-duplex LLM for realtime interaction. It unifies audio, video, and text in a single end-to-end architecture, watching, listening, and speaking simultaneously over continuous multimodal streams. The team reports that in end-to-end human evaluations it cuts conversational pacing issues roughly in half versus cascaded pipelines, with fewer interruptions, lower latency, and fewer false triggers. It supports joint audio-visual understanding, such as resolving homophones and ambiguous speech using the live scene and telling speakers apart in multi-party conversations, plus proactive interaction with tool use and natural conversational timing under noise and side talk. It fits realtime voice assistants, camera-based scenarios, and hands-free dialogue.

Seed-Music

Seed-Music is ByteDance Seed's suite of music generation systems for producing high-quality music with fine-grained style control. Built as a unified framework that adapts to musicians' evolving workflows, it combines auto-regressive language modeling with diffusion techniques: an LM performs controlled song and music generation with expressive vocals in multiple languages, conditioned on multi-modal inputs such as style descriptions, audio references, musical scores and voice prompts, while a diffusion model handles fine-grained note-level audio editing for post-production, including lyric and vocal melody edits. Its zero-shot voice conversion method needs only a 10-second singing or speech recording to bring a user's own voice into generated music, making it useful for songwriting demos, post-production touch-ups and interactive audio content.

UI-TARS-1.5

UI-TARS-1.5 is ByteDance Seed's open-source native GUI agent model, built on a vision-language model and trained with reinforcement learning to reason through its thoughts before taking action. Given a screenshot, it perceives desktop, browser, and mobile interfaces and executes actions such as clicking, typing, scrolling, and launching apps. Official results report state-of-the-art performance across GUI agent benchmarks, including 42.5 on OSWorld, 42.1 on Windows Agent Arena, 75.8 on Online-Mind2web, 64.2 on Android World, and 94.2 on ScreenSpot-V2, and it extends to game scenarios such as Minecraft and Poki games. The 7B variant is open-sourced on Hugging Face.

Gemini 2.5 Flash

Gemini 2.5 Flash is Google's workhorse thinking model, released in preview on April 17, 2025 in the Gemini API via Google AI Studio and Vertex AI. Google calls it its first fully hybrid reasoning model: developers can turn thinking on or off and set a thinking budget to balance quality, cost, and latency, and the model keeps 2.0 Flash-level speed even when thinking is off. It accepts multimodal input (text, images, audio, video), supports a 1M token context window, and is positioned on the performance-to-cost pareto frontier, making it a sensible default for cost-sensitive chat, summarization, agents, and interactive apps.

o3

o3 is OpenAI's most capable reasoning model, released broadly on April 16, 2025. It thinks deeply before answering and, for the first time in the o series, can agentically combine tools within its reasoning chain, including web search, Python for data analysis, image analysis, and image generation, deciding on its own when and how to use each. OpenAI reports state-of-the-art results on Codeforces, SWE-bench (without a custom scaffold), and MMMU, with clear gains over o1 in coding, math, science, and visual perception; when allowed to run Python on AIME 2025, o3 scored 98.4% pass@1 and 100% consensus@8. External expert reviewers found o3 makes 20% fewer major mistakes than o1 on hard real-world tasks. It fits complex analysis, multi-step workflows, and vision-heavy work such as interpreting charts, diagrams, and whiteboard photos, and is available in ChatGPT and through the API.