This SmarToken catalog tracks well-known global and Chinese AI models by where the teams are based, whether the release is proprietary or open, and application categories such as general LLMs, reasoning, image generation, video, music, audio and world-generation models. It was last updated on June 1, 2026.
This week's update adds Step's open reasoning model Step-3.7-Flash and MiniMax's planned-open-release agent model MiniMax M3.
Agent models, including models optimized for OpenClaw, are still an emerging category and are temporarily grouped with all-purpose multimodal or general models. Hybrid reasoning models are now mainstream, but many model families still release their general and thinking versions separately, so this catalog lists them as separate general and reasoning models. OCR models and 3D models are not yet treated as full categories in this version.
| Catalog note | How we classify it | Reader caveat |
|---|---|---|
| Agent models are still emerging. | They are temporarily grouped with all-purpose multimodal or general models. | The category may change as agent models mature. |
| Hybrid reasoning models dominate the market. | General and thinking versions are still listed separately when releases differ. | Compare the exact variant, not only the model family. |
| OCR and 3D models are not yet included. | They are not considered mature enough as standalone categories in this snapshot. | Future versions may add them. |
After two years of rapid development, the gap between Chinese AI models and models from outside China is shrinking, but overseas teams still keep an overall lead, especially in general-purpose models.
Google, OpenAI and Anthropic not only alternate in model performance leadership, but also still shape the main direction of the industry. There is also a practical warning for China-based companies: many overseas models require network workarounds to access from China, and their freer output style may create compliance risks in Chinese enterprise use.
The overseas proprietary section lists the main general and reasoning models from OpenAI, Google DeepMind, Anthropic, xAI and Mistral.
The table below keeps a catalog style. It is a dated market snapshot, so benchmark claims and release details need checking against current vendor pages before production use.
| Model or family | Category | Summary |
|---|---|---|
| GPT-5.5 / GPT-5.5 Instant | General | OpenAI's April 2026 general model shifts toward autonomous agents, native multimodality, 400K Codex context and stronger planning; Instant later became the default ChatGPT model with lower hallucination and more natural language. |
| Gemini 3.5 Flash | General | Google DeepMind's May 2026 model supports long-context work and focuses on fast iterative coding, advanced multimodal understanding, long workflows and multi-step tool use. |
| Claude Opus 4.8 / Sonnet 4.6 / Haiku 4.5 | General | Anthropic's line is especially strong for coding, agent work, multimodal reasoning and enterprise workflows; Chinese users should also consider policy and account-risk issues before relying on it. |
| Grok 4.1 | General | xAI's model follows a closed-latest, open-older-version strategy. The analysis says 4.1 improved emotional intelligence, writing human-likeness and hallucination rate. |
| Mistral 3 Medium / Mistral Large 2 | General | Mistral is treated as Europe's main AI lab, currently less active in frontier general-model competition and more focused on smaller models and niche innovations. |
| GPT-5.5 Thinking / GPT-5.5 Pro | Reasoning | Thinking is positioned for fast logical breakdowns and concise answers; Pro is the high-precision version for difficult tasks such as FrontierMath and Expert-SWE. |
| Gemini 3.1 Pro / Gemini 3 Deep Think | Reasoning | Major improvements include code-based animation, complex system synthesis, interactive design and difficult multimodal reasoning. |
| Claude Opus 4.8 Thinking | Reasoning | The key feature is Effort Control, allowing users to adjust thinking depth across low, high, extra and maximum settings. |
| Grok 4.1 Thinking / Magistral Medium v1.2 | Reasoning | Grok is xAI's thinking model; Magistral is a fast second-tier reasoning option with multimodal support. |
This catalog gives separate categories to image, video, music, audio and world-generation models because applications are now part of the model landscape. Here, world-generation models means systems that create or simulate interactive 3D environments.
This section follows the same catalog style: model name first, category second, and a short practical description. Some products are listed as applications rather than base models.
| Model or product | Category | Summary |
|---|---|---|
| Gemini 3.1 Flash Image / Gemini 3 Pro Image / Imagen 4 | Image | Google DeepMind's image line is stronger in text rendering, multilingual localization, complex instructions, character consistency and overall generation quality. |
| GPT-Image 2 / DALL·E 3 | Image | OpenAI's image model is a GPT-native visual system with thinking mode, internet-aware generation, self-checking and improved commercial production output. |
| Midjourney v7 / Flux 2 | Image | Midjourney remains an older image pioneer with slow updates; Flux 2 combines high-quality generation and editing in one architecture. |
| Veo 3.1 / Google Omni | Video | Veo improves audio, instruction following and realism; Google Omni combines Gemini reasoning with video editing and generation grounded in physical consistency. |
| Runway Gen-4.5 / Pika 2.5 / Luma AI / Stable Video Diffusion / Midjourney video | Video | These are major video-generation products, with limitations such as object disappearance, causal errors or local hardware requirements. |
| Lyrics 3 / Suno 5.0 | Music | Lyrics 3 supports text, image and video inputs and automatic lyric generation; Suno 5.0 improves studio-level quality, track separation and style control. |
| Stable Audio / MuseNet / V2A | Audio | These are audio-generation systems. V2A generates background audio for video from input footage and text prompts. |
| Genie 3 / World Labs Marble / World Labs RTFM | World-generation model | Genie 3 renders interactive virtual worlds in real time; World Labs focuses on persistent 3D environments and real-time generative worlds with spatial memory. |
The overseas open model section is shorter. It lists Mistral, Gemma, Phi, gpt-oss, Magistral, Muse Spark, Flux, Stable Diffusion and a few media models.
These open models form part of the wider international baseline, although Chinese open models occupy more of the practical comparison space in this catalog.
| Model or family | Category | Summary |
|---|---|---|
| Mistral Large 3 / Ministral 3 | General | Mistral Large 3 is listed as a 675B sparse MoE model with 256K context; Ministral 3 provides 3B, 8B and 14B edge models with multimodal ability. |
| Gemma 4 / Phi-4 | General | Gemma 4 emphasizes compute and memory efficiency, local deployment, function calling and multimodal understanding; Phi-4 is treated as a compact high-performing Microsoft model. |
| gpt-oss / Phi-reasoning / Magistral Small / Muse Spark | Reasoning | This group includes OpenAI's gpt-oss models, Microsoft's Phi reasoning line, Mistral's open Magistral Small and Meta's Muse Spark. |
| Flux 2 dev / Stable Diffusion | Image | Flux 2 dev is a strong open-weight image generation and editing model; Stable Diffusion remains the open local-deployment reference. |
| Hunyuan Video 1.5 / Wan 2.2 / audio and music models | Media | Open video, image, audio, music and world-generation entries appear later in the Chinese open model section. |
Chinese large-language models were initially rushed into the market, but after more than a year of development the gap with leading overseas teams has narrowed, especially in music, image generation, video generation and reasoning models.
Many early Chinese LLMs were launched quickly to catch the market, but the strongest Chinese players have since become serious contenders in several categories.
The Chinese proprietary section starts with consumer and platform model families such as Doubao, Qwen, GLM, Kimi, Tencent Yuanbao, Step, ERNIE and MiMo.
The table below follows the same category order used throughout this catalog and keeps the main practical distinctions clear.
| Model or family | Category | Summary |
|---|---|---|
| Doubao Seed 2.0 | General | ByteDance's general model application line. Seed 2.0 improves visual reasoning, temporal and motion perception, instruction following and complex agent-task performance. |
| Qwen3.6 Plus / Qwen3.5 Omni | General / All-purpose multimodal | Alibaba's Qwen line is application-oriented and broad. Qwen3.6 Plus emphasizes agentic coding and 1M context; Qwen3.5 Omni supports long audio/video input, real-time interaction and multilingual speech. |
| GLM-5V-Turbo / GLM-5-Turbo | Agent / General | Zhipu AI's models are described as agent-oriented, with GLM-5V-Turbo focused on visual programming and GLM-5-Turbo focused on tool use, instruction following and long-chain execution. |
| Kimi | General | Moonshot's Kimi was once China's long-document reading king, though its market presence became quieter after entering 2025. |
| Tencent Yuanbao / Hunyuan HY2.0 | General | Tencent's Yuanbao, formerly Hunyuan, is a MoE model family with improved reasoning, coding, agent and instruction-following ability, though with weaker visibility after integrating DeepSeek-R1. |
| ERNIE 5.1 | General | Baidu's model is an ultra-sparse MoE with much lower pretraining cost than comparable models and a four-stage post-training pipeline. |
| MiMo-V2.5-Pro / MiMo-V2.5 | Agent / General | Xiaomi's MiMo models emphasize long-cycle autonomous planning, multimodal action and 1M context in a unified model. |
| Doubao Seed 2.0 Pro / Qwen3.7 Max / Step R1 v-mini / X1 Turbo / HY2.0 Think | Reasoning | These are major Chinese closed reasoning models, with Qwen3.7 Max especially framed around long autonomous execution and cross-framework generalization. |
The Chinese application section is broad. It includes Seedance, Kling, Hailuo, Qingying, Wan, MiniMax Music, Mureka, StepAudio, Speech, Qwen TTS, Kling Image, Seedream, Wan image models and Happy Oyster.
This part is closer to an application catalog than a base-model comparison, so the table keeps a product-oriented structure.
| Model or product | Category | Summary |
|---|---|---|
| Seedance 2.0 / Kling 3.0 Omni / Kling 3.0 | Video | Seedance 2.0 improves reference control and editing; Kling 3.0 Omni focuses on multimodal consistency, while Kling 3.0 focuses on professional video generation and narrative control. |
| Hailuo 2.3 / Qingying / Wan 2.7 Video / HappyHorse | Video | MiniMax Hailuo improves body movement, facial expression and physics; Zhipu Qingying supports cinematic parameters; Wan 2.7 covers text-to-video, image-to-video, reference video and editing; HappyHorse is listed as a placeholder dark-horse entry. |
| MiniMax Music 2.6 / Mureka V7 | Music | MiniMax Music improves Chinese-style music, epic low-end sound, lo-fi or indie-folk looseness and cover workflows; Mureka V7 uses MusicCoT to plan the musical structure before filling content. |
| StepAudio 2.5 TTS / MiniMax Speech 2.6 / Qwen3-TTS | Audio | StepAudio focuses on contextual TTS and emotional delivery; MiniMax Speech reduces latency and improves imperfect voice cloning; Qwen3-TTS emphasizes Chinese and English stability and many expressive voices. |
| Kling Image 3.0 Omni / Seedream 4.5 / Wan 2.7 Image | Image | Kling Image focuses on cinematic narrative visuals; Seedream is optimized for Chinese semantics and advanced editing instructions; Wan 2.7 strengthens virtual-avatar design, color palette control and long text/formula rendering. |
| Happy Oyster | World-generation model | Alibaba's token team is running a small internal test for a world-generation model, with limited access at the time of this update. |
The Chinese open model section is the most detailed part of this catalog. It lists DeepSeek, Qwen, GLM, Kimi, MiniMax, ERNIE, MiMo, Hunyuan and Step.
Open Chinese models are a major part of the market. This section pays close attention to total parameters, active parameters, context length, agent ability, multimodal support and deployment efficiency.
| Model or family | Category | Summary |
|---|---|---|
| DeepSeek V4 Pro / V4 Flash | General | DeepSeek V4 Pro is a 1.6T total, 49B active hybrid reasoning flagship with 1M context, compressed sparse attention and stronger agent ability; V4 Flash is smaller, faster and cheaper with similar simple-task ability but weaker hard-task performance. |
| Qwen3.6-27B / Qwen3.6-Flash / Qwen3.5 / Qwen3-VL / Qwen3-Omni / Qwen3-Next | General / Multimodal | Qwen's open line is broad, from 27B dense multimodal models to 35B-A3B sparse models, 397B-A17B Qwen3.5, VL, Omni and new architecture routes. |
| GLM-5.1 / GLM-4.6V | General | Zhipu's open models emphasize long work sessions, planning-execution-iteration loops, agentic coding, multimodal output and rich content creation. |
| Kimi K2.6 | General | Moonshot's open model is a 1T total, 32B active model with 256K context, stronger general agent, code and visual understanding ability than K2.5. |
| MiniMax M3 | Agent / General | MiniMax M3 is a 1M-context, native multimodal agent model with coding, agent collaboration and computer-use ability, using MiniMax Sparse Attention. |
| ERNIE 4.5 / MiMo V2 Flash / Hy3-preview / Step3-VL | General / Multimodal | Baidu, Xiaomi, Tencent Hunyuan and Step open models are important Chinese entries, each with different MoE, multimodal, reasoning or agent strengths. |
| DeepSeek R1 / Qwen thinking models / GLM thinking models / Seed OSS / Hunyuan A13B / Kimi thinking models / Step-3.7-Flash / MiniMax M2.1 | Reasoning | The reasoning section gathers open Chinese thinking models, with Step-3.7-Flash newly highlighted for production agent workflows, multimodal understanding, search and compatibility with mainstream agent frameworks. |
The final catalog section lists Chinese open models for video, image, audio, music and world generation.
This is where the catalog moves beyond LLMs into generation systems and world-building tools. The category split matters because these products should not be judged only by text-model benchmarks.
| Model or family | Category | Summary |
|---|---|---|
| Hunyuan Video 1.5 / Wan 2.2 | Video | Hunyuan Video 1.5 is a smaller open video model with improved motion, aesthetics and preference alignment; Wan 2.2 supports 720P, 24fps text-to-video and image-to-video generation on consumer GPUs. |
| Z-Image / Qwen-Image | Image | Z-Image is a 6B scalable single-stream DiT image model with Turbo, Base and Edit variants; Qwen-Image is a 20B image foundation model strong in complex text rendering and precise image editing. |
| Hunyuan Image / GLM-Image / CogView | Image | Hunyuan Image is a large open image MoE model; GLM-Image combines autoregressive and diffusion decoding; CogView remains Zhipu's bilingual DiT image model. |
| HunyuanVideo-Foley | Audio | Tencent's open 3B audio-generation model creates layered sound effects by understanding video frames and text descriptions together. |
| ACE-Step | Music | StepFun and ACE Studio's 3.5B open music model is listed as the open music entry. |
| HY-World 2.0 | World-generation model | Tencent Hunyuan's world-generation model accepts text, single-view images, multi-view images and video, then generates 3D world representations such as meshes or 3D Gaussian splats. |
This catalog includes a long update log covering 2026 and 2025. The log is part of its value because it shows how quickly model categories and release details change.
The latest update entries add Claude Opus 4.8, Qwen3.7-Max, Gemini 3.5 Flash, Gemini Omni, ERNIE 5.1, GPT-5.5 Instant, DeepSeek V4, GPT-Image 2, Hy3-preview, MiMo, StepAudio, Kimi K2.6, Qwen3.6, GLM-5V-Turbo, Wan 2.7, Gemma 4 and other model families. Older 2025 entries track updates to DeepSeek, Qwen, GLM, MiniMax, Kimi, Hunyuan, Claude, Gemini, Grok, Flux, Runway, Kling, Wan and many others.
| Update period | Examples from the update log | Meaning |
|---|---|---|
| June 2026 | Step-3.7-Flash and MiniMax M3. | The latest update emphasizes reasoning and agent models. |
| May 2026 | Claude Opus 4.8, Qwen3.7-Max, Gemini 3.5 Flash, Gemini Omni, ERNIE 5.1 and GPT-5.5 Instant. | Top overseas and Chinese labs keep refreshing general, reasoning and multimodal models. |
| April 2026 | DeepSeek V4, GPT-5.5, GPT-Image 2, Hy3-preview, MiMo, StepAudio, Kimi K2.6, Qwen3.6 and new world-model entries. | The April log is one of the densest updates in the catalog. |
| 2025 updates | DeepSeek, Qwen, GLM, MiniMax, Kimi, Hunyuan, Claude, Gemini, Grok, Flux, Runway, Kling, Wan and others. | The catalog tracks model changes month by month rather than treating the market as static. |
Mistake
Treating one article as a final ranking
Why it hurts
Model releases, pricing, quotas and benchmark positions can change quickly.
Better move
Use the analysis as a shortlist, then run current checks against your own workload.
Mistake
Choosing by brand instead of task
Why it hurts
A strong chat model may still be weak for long documents, coding agents, multimodal work or low-latency routes.
Better move
Define the job first, then compare models with prompts, files or media that match that job.
Mistake
Copying claims without a current verification check
Why it hurts
Benchmark numbers, context windows, API names and prices may be dated or provider-specific.
Better move
Confirm high-impact details against official docs, model cards or live provider pages.
Use this page to understand the model family, the evaluation angle and the current conversation around it. Then choose one or two realistic prompts, documents or media tasks and test whether the model behaves well in your own workflow.
These questions reflect recurring reader concerns around Chinese model knowledge, evaluation and fast-moving model releases.
This SmarToken catalog tracks well-known global and Chinese AI models by where the teams are based, whether the release is proprietary or open, and application categories such as general LLMs, reasoning, image generation, video, music, audio and world-generation models. It was last updated on June 1, 2026.
Use it as market and product context, then verify technical claims, pricing, quotas and release details against official pages or your own tests before making a decision.
The video gives a fast visual summary of the model story, while the written page carries the caveats, comparisons and practical checks.
SmarToken tracks public model releases, technical reports, product announcements and market signals to keep this catalog useful.
Technical claims need to be treated as dated unless they are confirmed by current official model cards, technical reports or provider announcements.
Pricing, quota, availability and benchmark details can change after the review date, so production decisions should use current vendor pages and direct workload tests.
