# Chinese AI model names, decoded (markdown mirror)

Source: https://aiapi-pro.com/blog/chinese-ai-model-naming-guide-2026

A decoder for 111 model ids from 14 providers. Catalog on 2026-10-22.

## The four layers

| Layer | Example | Question |
|---|---|---|
| Vendor / family | `glm`, `kimi`, `doubao`, `qwen` | Who made it? |
| Product line | `seedream`, `seedance`, `hy-3d`, `hy-mt` | What is it for? |
| Generation | `v4`, `5.3`, `k2.7`, `m2.5` | How new? |
| Variant suffix | `-flash`, `-pro`, `-turbo`, `-code` | Which trade-off? |

Vendors order these differently: ByteDance puts the product line inside the name (`doubao-seedance-2.5`), Alibaba mixes an internal generation into the prefix (`qwen3.7-max`), Tencent sometimes drops the vendor (`hy3`).

## Suffix dictionary

| Suffix | Means | Price signal | Example |
|---|---|---|---|
| `-flash` | cheap high-throughput tier | lowest paid | `glm-5.3-flash` $0.08/1M in |
| `-flashx` | branded flash sub-variant | between flash and full | `glm-5.3-flashx` $0.12/1M in |
| `-lite` | smallest tier | lowest | `doubao-seed-2.0-lite` $0.075/1M in |
| `-pro` | quality tier | higher than flash | `deepseek-v4-pro` $0.57/1M in |
| `-max` / `-plus` | flagship | highest in family | `qwen3.8-max`, `qwen-mt-plus` |
| `-turbo` | speed-optimised | family-dependent | `glm-5-turbo` |
| `-highspeed` | throughput-optimised | often more expensive | `kimi-k2.7-code-highspeed` $2.07/1M in |
| `-code` | tuned for code | tag, not tier | `doubao-seed-2.0-code` |
| `-vision` / `-vl` | accepts images | tag | `qwen3-vl-plus` |
| `-thinking` | exposes reasoning | tag | `glm-4.1v-thinking-flash` |
| `-omni` | multi-modal in/out | tag | `qwen3-omni-flash` |
| `-preview` | pre-release snapshot | no guarantee | `hy4-preview` |
| `-mt` | machine translation | cheapest per token | `hy-mt2-lite` $0.043/1M in |
| `-asr` / `-tts` | speech recognition / synthesis | billed differently | `qwen3-tts-flash` |
| `-t2i` | text-to-image | per image | `wan2.2-t2i-flash` |
| `-3d` | 3D generation | per generation | `hy-3d-motion` |
| `-embedding` | embeddings | per 1M, no output price | `qwen3.7-text-embedding` |

## Vendor cheat sheet

| Vendor | Chat/LLM | Image | Video | Other |
|---|---|---|---|---|
| DeepSeek | `deepseek-v4-flash`, `-pro`, `v4.1-flash` | - | - | `-vision` |
| Zhipu AI | `glm-5.x` (+ flash/flashx/turbo) | `cogview-3-flash` | `cogvideox-flash` | `glm-4.6v` |
| Alibaba Cloud | `qwen3.7-max`, `qwen-plus`, `qwen3-coder-flash` | `wan2.x-t2i-*` | - | `qwen-mt-plus`, `qwen3-tts-flash`, OCR, embeddings |
| Moonshot | `kimi-k3`, `kimi-k2.x` (+ `-code`) | - | - | `-highspeed` |
| ByteDance | `doubao-seed-2.x` (lite/pro/code/turbo) | `seedream-*` | `seedance-*` | - |
| Tencent Hunyuan | `hy3`, `hy4-preview`, `hunyuan-*` | `hy-image-v3` | `hy-video-v1.5` | `hy-mt2-*`, `hy-3d-*`, ASR |
| MiniMax | `minimax-m2.5/m2.7/m3`, `minimax-text-01` | - | `minimax-video-h3` | `minimax-speech-*`, `minimax-music-*` |
| Xiaomi | `mimo-v2.5-pro`, `mimo-v2.6-flash` | - | - | - |
| Kuaishou | - | - | `kling-video-v3` | `-omni`, `-turbo` |
| Others | `step-5-preview` | `wand-vega-image-*`, `vidu-image-q2` | `pixverse-video-*`, `vidu-video-q3` | `tripo-3d-*`, Kinfra embeddings |

## Four traps

1. **"Flash" is a price tier, not a free plan.** GLM-5.3-Flash $0.08/1M; GLM-5.3 $1.25. A few flash models are free to call on the free tier — a plan decision, not a naming rule.
2. **Family-internal comparisons do not transfer.** Doubao-Seed-2.0-Pro $0.40/1M vs Kimi K3 $2.90/1M: a "Pro" at a seventh of another vendor's flagship.
3. **Aliases appear twice.** `doubao-seedream-5.0` and `seedream-image-v5.0-pro` are the same ByteDance image line under two conventions.
4. **"Preview" is a promise about nothing.** Preview snapshots can change or disappear: fine for evaluation, risky in production.

## How to pick without memorising this

1. Resolve the live list: `GET /v1/models` or the catalog page.
2. Filter by modality/context on the cheapest-per-task page.
3. Estimate spend with the cost calculator.
4. Start on the free tier for first impressions.

## FAQ

**What does "flash" mean in a Chinese model name?**
Flash marks the cheap, high-throughput tier of a family - not literally free. GLM-5.3-Flash is $0.08 per 1M input tokens while GLM-5.3 is $1.25. Some flash models are genuinely free to call on NovAI's free tier.

**What is the difference between Seedream, Seedance and Doubao?**
They are ByteDance product lines, not model generations: Doubao-Seed-* is the chat/LLM line, Seedream is image generation, Seedance is video generation.

**Why do some models have two ids?**
Aliases. doubao-seedream-5.0 and seedream-image-v5.0-pro refer to the same ByteDance image line under different naming conventions.

**Does "pro" always cost more than "flash"?**
Within one family, almost always. Across families it says nothing. Doubao-Seed-2.0-Pro is $0.40 per 1M input, cheaper than Kimi K3 at $2.90.