New AI Models

Qwen3.7-Max-2026-05-17

Qwen3.7-Max-2026-05-17 is the dated snapshot of the Qwen3.7-Max preview, the largest and most capable model in the Qwen3.7 series. It supports thinking mode only and offers a text-only interface for experimentation, optimized for general-purpose conversational use cases such as knowledge-based question answering, instruction following, and creative writing. Pricing is 12 CNY per million input tokens and 36 CNY per million output tokens, with up to 1M input tokens per request.

Qwen3.7-Max-Preview

Qwen3.7-Max-Preview is the preview edition of the largest and most capable Max model in Alibaba's Qwen3.7 series. It supports thinking mode only and is currently available as a text-only model, optimized for general conversational use cases such as knowledge Q&A, instruction following, and creative writing. It offers a 1,000,000-token context window and is available on Alibaba Cloud Model Studio (Bailian) at 12 CNY per million input tokens and 36 CNY per million output tokens.

Antigravity 2.0

Google Antigravity 2.0 is a standalone, agent-first desktop app for macOS, Linux, and Windows, powered by the latest Gemini models. New capabilities include dynamic subagents, asynchronous task management, JSON hooks, and cron-based Scheduled Tasks, with feedback given directly on agent artifacts. Alongside a CLI, SDK, and API, it targets both software development and broader knowledge work, and is available from the official download page.

Gemini Omni

Gemini Omni is Google's natively multimodal generative model family for creating high-quality video from any combination of image, audio, video, and text input. It supports multi-turn conversational editing in natural language, keeping characters and physics consistent across instructions, and grounds its output in Gemini's world knowledge. The first model in the family, Gemini Omni Flash, is available to Google AI Plus, Pro, and Ultra subscribers through the Gemini app and Google Flow, and every generated video carries a SynthID watermark.

Gemini 3.5

Gemini 3.5 is Google's latest model family for agentic workflows, debuting with Gemini 3.5 Flash and with 3.5 Pro to follow. 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks such as Terminal-Bench 2.1, MCP Atlas and GDPval-AA, delivers output tokens 4 times faster than other frontier models, and often runs at less than half the cost. It targets long-horizon agentic tasks and is available via the Gemini app, Search AI Mode, Google AI Studio and Gemini Enterprise.

GPT-5.5-Cyber

GPT-5.5-Cyber is OpenAI's cybersecurity model in limited preview, available to defenders responsible for securing critical infrastructure. Specially trained to be more permissive on authorized security tasks than GPT-5.5, it pairs that flexibility with stronger verification and account-level controls for workflows such as authorized red teaming, penetration testing, and controlled validation. Access is provided through the Trusted Access for Cyber (TAC) program, with individuals verifying identity at chatgpt.com/cyber.

new realtime voice models

OpenAI ships three new audio models in the Realtime API. GPT-Realtime-2 brings GPT-5-class reasoning to live voice, with parallel tool calls, a 128K context window, and adjustable reasoning effort for balancing latency against depth. GPT-Realtime-Translate translates speech from 70+ input languages into 13 output languages in real time, while GPT-Realtime-Whisper streams low-latency speech-to-text for captions and meeting notes. GPT-Realtime-2 is priced at $32 per 1M audio input tokens; the Translate and Whisper models cost $0.034 and $0.017 per minute.

Mistral Medium 3.5

Mistral Medium 3.5 is Mistral AI's first flagship merged model, a dense 128B model with a 256k context window that combines instruction-following, reasoning, and coding in a single set of open weights. It scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and Qwen3.5 397B A17B, and self-hosts on as few as four GPUs. Now the default model in Mistral Vibe and Le Chat, it targets vibe coding, remote cloud agents, and long-horizon productivity work, priced at $1.5 per million input and $7.5 per million output tokens via API.

hunyuan-image-v3.0-v1.0.5

HunyuanImage-3.0 is Tencent Hunyuan's native multimodal model for image generation, built on a unified autoregressive architecture with Mixture of Experts, 80B total parameters and 13B activated per token. The official platform refreshed the image 3.0 line in late April 2026 with multi-turn interaction, more precise text rendering, and stronger photorealistic aesthetics. Weights are open under the Tencent Hunyuan Community License, suited to text-to-image work and text-heavy visuals like posters.

MiMo-V2.5-Pro

MiMo-V2.5-Pro is Xiaomi's most capable model to date, a 1.02T-parameter Mixture-of-Experts model with 42B active parameters and a 1M-token context window. It targets agentic work and long-horizon software engineering, sustaining workflows of more than a thousand tool calls, and scores 64% Pass^3 on ClawEval. The weights are fully open-sourced under a permissive license on Hugging Face, with deployment guides for SGLang and vLLM.

HappyHorse-1.0-R2V

HappyHorse-1.0-R2V is a reference-to-video model available on Alibaba Cloud's Bailian platform. It accepts up to nine reference images and blends the subjects from those images into a coherent video guided by a text prompt, with stable subject and scene reference that keeps the creative intent intact. Videos run 3 to 15 seconds at up to 1080P, with multiple aspect ratios such as 16:9 and 9:16.

HappyHorse-1.0-Video-Edit

HappyHorse-1.0-Video-Edit is a video editing model offered on Alibaba Cloud Model Studio (Bailian). It follows natural-language instructions to edit an existing video, handling tasks such as style transfer and partial replacement, and accepts up to 5 reference images for local or global edits while accurately reproducing the original motion. Outputs are MP4 clips of 3 to 15 seconds at 720P or 1080P, with an option to keep the source audio.

Wan2.7-Image-To-Video-2026-04-25

Wan2.7-Image-To-Video is Alibaba Cloud's image-to-video model in Model Studio, and this 2026-04-25 snapshot is the recommended pinned version. It supports first-frame, first-and-last-frame, and video continuation generation with audio, outputting 720P or 1080P MP4 at 30fps for 2 to 15 seconds. Acting quality is upgraded across the board, from nuanced emotional scenes to intense action, with more cinematic shot transitions.

Wan2.7-Text-To-Video-2026-04-25

Wan2.7 Text-to-Video (2026-04-25 snapshot) is Alibaba's Wan video generation model that creates cinematic multi-shot videos with sound from text prompts, and it also accepts audio input for synchronized sound and picture. Its acting capability is upgraded across the board, covering nuanced emotional scenes, intense action, and more dramatic, well-paced shot transitions. It supports 720P and 1080P output at 30fps in MP4 format, with durations from 2 to 15 seconds.

Seed3D 2.0

Seed3D 2.0 is ByteDance Seed's new-generation 3D generation model built for production-ready image-to-3D asset creation. It pairs two-stage DiT geometry generation with unified PBR texture synthesis, and adds part-level generation and articulation modeling that exports URDF content compatible with Isaac Sim. In pairwise blind evaluation against six mainstream models across about 200 test cases, it achieved SOTA results in both geometry and texture generation, with a texture preference rate above 69%. The API is available on Volcano Engine as Doubao-Seed3D-2.0.

DeepSeek-V4-Flash-0423

DeepSeek-V4-Flash is the fast, cost-effective model in DeepSeek's open-source V4 Preview lineup, built on a MoE architecture with 284B total and 13B active parameters. Its reasoning closely approaches V4-Pro, and it performs on par with V4-Pro on simple agent tasks while offering faster responses and highly cost-effective API pricing. It supports a 1M context window with Thinking and Non-Thinking modes, and integrates with coding agents such as Claude Code and OpenCode.

DeepSeek-V4-Flash-Max

DeepSeek rolled out its V4 series in April 2026 with the V4 Preview, introducing the flagship V4-Pro alongside the fast-tier V4-Flash. The Flash tier is a MoE model with 284B total and 13B active parameters, and a 1M context window is standard across the lineup. Reasoning closely approaches V4-Pro and matches it on simple agent tasks, with faster responses and more cost-effective API pricing for high-volume chat, agent, and coding workloads.

DeepSeek-V4-Pro-Max

DeepSeek-V4-Pro-Max is the flagship open-source text model in DeepSeek's V4 lineup, built with 1.6T total and 49B active parameters and a default 1M context window backed by token-wise compression and DSA sparse attention. Official benchmarks report open-source SOTA in agentic coding, with math, STEM, and coding results ahead of all current open models and world knowledge second only to Gemini-3.1-Pro. It supports Thinking and Non-Thinking modes and fits agentic coding and long-context reasoning workloads.

GPT-5.5

GPT-5.5 is OpenAI's most intelligent model to date, released in April 2026 for agentic work in coding, research, and computer use. Available as GPT-5.5 Thinking and GPT-5.5 Pro in ChatGPT, it matches GPT-5.4 latency while scoring 82.7% on Terminal-Bench 2.0 and 78.7% on OSWorld-Verified with lower token usage. API pricing is $5 per million input tokens and $30 per million output tokens with a 1M context window, fitting agentic coding, knowledge work, and scientific research.

Qwen-Image-2.0-Pro-2026-04-22

Qwen-Image-2.0-Pro-2026-04-22 is a dated snapshot of Qwen-Image-2.0-Pro, the full-capability image generation and editing model in the Qwen-Image-2.0 series, currently matching the rolling qwen-image-2.0-pro endpoint. It offers professional text rendering with 1k token prompt support, finer photorealistic texture, and stronger semantic adherence, with the best text rendering and realism in the 2.0 series. As a fixed snapshot, it suits production workloads that need stable output. Total output pixels range from 512*512 to 2048*2048, with up to 6 images per call.