Models

Vidu Q4 Preview

Next-gen flagship video model preview from Shengshu Technology, focused on expressive character performance, cinematic camera work, and complex VFX . Supports up to 15 reference images (characters, wardrobe, props, scenes) and 3 audio references for voice consistency . Generates up to 16-second clips at 540p–4K with 10-bit color depth

Seedance 2.5

ByteDance's latest flagship video generation model, built for longer-form storytelling and production-ready output. Generates up to 30 seconds of continuous, cinematic video with native audio sync in a single pass . Accepts up to 50 multimodal references (images, videos, audio, character sheets, storyboards) for precise scene, character, and motion consistency . Features localized region editing to fix specific areas without full regeneration……

Wan 3.0

Wan3.0-Video is an all-in-one video generation model unified support for multiple creative capabilities, including reference, editing, replication, and driving. It generates videos up to 30 seconds with omni-modal reference, and can parse files, web pages and complex images. With production-grade character consistency and lifelike visuals and sound, it delivers an immersive audiovisual experience.

Seedance 2.0 Mini

Lightweight, cost-efficient video model from ByteDance, optimized for speed and high-volume content creation. Supports text-to-video, image-to-video, and reference-based generation with up to 12 references (6 images, 3 audio, 3 video). Delivers faster generation and lower credit consumption than Seedance 2.0, with strong motion quality and character consistency. Ideal for social media content, product videos, AI short dramas, and rapid creative iteration

GLM-5.3

Zhipu AI's flagship text model optimized for complex software engineering and long-horizon Agent tasks. Features a 1M-token context window with mandatory reasoning (3 levels: low/high/max). Coding capability improved 50% over GLM-5.2 on Z.ai Code Bench; scores SOTA on Terminal Bench 3.0. Excels in cybersecurity tasks (cyber vulnerability discovery)

Seedance 2.0

Generate videos from reference images, videos, and audio; edit videos; extend videos; generate videos from start and end frames

Wan 3.0 Prime

Wan3.0-Video-Prime is the high-speed video generation model of Wan3.0, with capabilities aligned to the standard version of Wan3.0-Video. It supports four-modal all-in-one reference and generates videos of up to 30 seconds, delivering an immersive audiovisual experience with significantly faster end-to-end generation.

Seedance 2.0 Fast

Generate videos with reference to images/videos/audio, edit videos, extend videos, generate videos from first and last frames

Kling 3.0 Omni

Flagship unified multimodal model integrating text-to-video, image-to-video, and reference-based generation. Supports up to 15-second cinematic clips with native synchronized audio (dialogue, SFX, BGM). Enables multi-shot control (up to 6 shots) and consistent subject/character preservation across scenes. Pro mode outputs 1080p with enhanced motion realism

kling v3

Next-generation video generation model offering Standard and Pro tiers. Generates 3–15 second clips at up to 1080p resolution from text or image inputs. Features first-frame and last-frame control for precise scene composition. Supports 16:9, 9:16, and 1:1 aspect ratios. Native audio generation available as an optional feature

GLM-5.3 Flash

Z.ai's native multimodal model (320B-A18B) with text, image, video, and file input. Features hybrid sparse + linear attention, reducing attention compute 3x and KV cache 4.4x vs. GLM-5.3. Scores 57 on Artificial Analysis Intelligence Index—matching Claude Opus 4.8—with 1M context and 128K max output. Excels in agentic coding, vision-based Computer Use, and CAD reproduction

Kling Video O1

World's first unified multimodal video model built on MVL (Multi-modal Visual Language) architecture. Accepts multimodal inputs—text, images, videos, and elements—for all-in-one creation and editing. Supports reference-based generation, start/end frame interpolation, video in/outpainting, stylization, and multi-subject consistency. Generates 3–10s clips with up to 7 reference images

MiniMax-H3-Max

High-speed video model post-trained by fal.ai on MiniMax H3, optimized for low-latency generation. Ranks #1 in image-to-video (Elo 1,204) and #3 in text-to-video with audio (Elo 1,235) on Artificial Analysis. Generates 5-second 768p video in under 3 seconds (35x throughput vs. H3), enabling real-time streaming, live AI broadcasting, and instant content creation. Supports text-to-video and image-to-video (first/last frame), 5–15s clips, 480p/768p, and native stereo audio

Qwen3.8 Flash

Alibaba's cost-efficient multimodal reasoning model. Supports text, image, and video inputs with text output. Features a native 1M-token context window for long documents, codebases, and agentic workflows. Excels in coding assistance, desktop interaction, chart analysis, and long-video understanding. Compatible with OpenAI/Anthropic protocols for seamless integration

GLM-5.2

Flagship text model purpose-built for long-horizon agentic workflows. Features a 1M context window supporting project-level engineering in a single session. Excels at autonomous coding: can complete development, testing, and multi-platform deployment from a single prompt. Top open-weight model per Artificial Analysis; #1 globally on Code Arena. MIT-licensed and Day-0 optimized for domestic AI chips

MiniMax H3

Open-weight general-purpose multimodal video model. Unifies text, image, video, and audio understanding in a single context window, generating up to 2K resolution, 15-second clips with native stereo audio at 24fps. Supports multimodal reference inputs: up to 9 images, 3 videos, and 3 audio clips (12 total references) per generation. Features first-frame, last-frame, and full reference modes with conversational editing capabilities.

Qwen3 Max

Compared with the September 23, 2025 version, the newly upgraded Qwen-3 Max seamlessly integrates thinking and non-thinking modes, bringing an all-round obvious performance boost. Its thinking mode supports web search, web content extraction and code interpreter. It can conduct in-depth logical reasoning and call external tools to solve intricate problems more precisely

GLM-5

GLM-5 is Z.ai’s flagship open-source foundation model engineered for complex systems design and long-horizon agent workflows. Built for expert developers, it delivers production-grade performance on large-scale programming tasks, rivaling leading closed-source models. With advanced agentic planning, deep backend reasoning, and iterative self-correction, GLM-5 moves beyond code generation to full-system construction and autonomous execution.

Seedream 5.0

Supports text, single-image and multi-image inputs, and enables the generation of image sets

DeepSeek-V4.1-Flash

Lightweight flagship of DeepSeek's new Causal Encoder-Decoder family. 552B-parameter MoE with asymmetric activation—8B for input, 16B for output—delivering flagship-level intelligence at significantly lower inference cost . Native multimodal vision understanding. 1M-token context window with 384K max output. KV cache compressed to 1/4 HBM and 1/8 SSD of previous generation . Outperforms V4-Pro on Terminal-Bench 2.1 (90.6), DeepSWE v1.1 (74.2), and CyberGym (88.1)

MiniMax M3

Flagship multimodal foundation model supporting text, image, and video inputs with text output. Features a 1M-token context window via MiniMax Sparse Attention (MSA), cutting per-token compute to ~1/20 of previous gen at full context. Excels at long-horizon agentic work, coding, and tool use. Native multimodal training from step zero ensures deep semantic alignment. Scores 59.0% on SWE-Bench Pro and 83.5 on BrowseComp, surpassing Opus 4.7

mimo-v2.6-pro

Xiaomi's flagship open-source omni-modal model (1T params) with 1M context. Scores 46 on Artificial Analysis Intelligence Index—top open-weight model, surpassing Kimi K3 and Qwen3.8 Max. Excels in agentic coding (DeepSWE v1.1: 72.57), 3D spatial reasoning, and CUA.

mimo-v2.6-flash

Efficient omni-modal model (309B total, 15B activated) with hybrid attention and 1M context. Outperforms previous MiMo-V2.5-Pro across benchmarks; DeepSWE v1.1 score improved to 65.68 via scaled RL training. Optimized for agentic workflows, coding, and visual tasks

Qwen3 Coder Plus

Powered by Qwen3, this is a powerful Coding Agent that excels in tool calling and environment interaction to achieve autonomous programming. It combines outstanding coding proficiency with versatile general-purpose abilities.