Chinese AI Models Overview [September 2026]
A permanent overview of major Chinese models — DeepSeek, Qwen (Alibaba), GLM (Zhipu/Z.ai), Kimi (Moonshot), MiniMax, StepFun and more. We prioritize primary sources (official announcements & pricing pages); where official figures are unavailable, third-party reporting/evaluations are clearly attributed. No speculation or rumors. See Pricing Comparison and the DeepSeek guide.
Why track Chinese AI
Open weights and aggressive pricing give these models outsized reach in both API and self-hosted use.
- Origin of the "DeepSeek Shock" — DeepSeek R1 (Jan 2025) shipped o1-class reasoning at low cost under an MIT license, resetting industry pricing and release strategy.
- The open-weight front line — Qwen, GLM, Kimi, MiniMax and DeepSeek have all released flagship-class open models, many under MIT / Apache 2.0-style licenses.
- Price competitiveness — DeepSeek's peak/off-peak model and low-cost Qwen/GLM flagships can cost a fraction of comparable Western models.
- Caveats — Self-reported benchmarks can diverge from independent evaluations. Check data-sovereignty, regional and ToS considerations too.
Major labs & flagships
DeepSeek (深度求索)
China's most prominent lab after the R1 "DeepSeek Shock." Current: V4.1 Flash (released 2026-09-10; 1M context, 384K output, native image input). V4 Flash and Vision Exp were consolidated into V4.1 Flash. V4 Pro keeps running after 9/14 (the 9/14 routing notice was withdrawn). MIT license. Peak/off-peak pricing since 2026-08-16. Guide · What changes · Pricing.
Qwen (通义千问 / Alibaba Cloud)
Alibaba's open-weight family. Flagship Qwen3.8-Max (GA 2026-08-03) is a 2.4T-parameter MoE with 1M context. Low-cost Qwen3.8-Flash (2026-08-26) is a 125B/6B active MoE previewing the Qwen4 architecture (open-weight Flash-Next on HF/ModelScope; check each model card for license terms). Qwen3.8-Flash details.
GLM (Zhipu AI / Z.ai)
Tsinghua-affiliated lab. GLM-5.2 (2026-06-16) is a 744B MoE under MIT; GLM-5.3 (2026-08-14) ships at the same price. GLM-5.3-Flash (2026-08-26) is a 320B/18B MoE, natively multimodal, at list $0.15/$0.50. First foundation-model AI to IPO (Hong Kong, Jan 2026). GLM 5.3 Flash details.
Kimi (Moonshot AI / 月之暗面)
Beijing startup. Kimi K3 (2026-07-16) is a 2.8T-parameter model with native vision and a 1M-token context, priced at $3.00 input / $15.00 output, with cached input at $0.30 (Moonshot direct API) — Claude Sonnet territory rather than the budget tier. It is also served through third-party hosts such as OpenRouter, and became available on Amazon Bedrock in September 2026 (AWS blog, 2026-09-18). K2.6 (Apr 2026) is reported to match GPT-5.5 on SWE-bench Pro.
MiniMax
M3 (May/June 2026) is a multimodal MoE with 1M context. M2.7 open-sourced 2026-04-12. Low-cost, agent-oriented models.
StepFun (阶跃星辰)
A Chinese AI startup. Its official documentation splits pricing by model category, and one of those categories is “Multimodal Reasoning Models”, where as of 2026-10-13 it lists step-5-preview ($1.00 / $2.70) and step-3.7-flash ($0.20 / $1.15), with cached input at $0.10 and $0.04 respectively. The wider catalogue spans text-only, vision, speech and image generation — for speech, stepaudio-2.5-realtime ($1.50/$10.00) and step-audio-r1.5 ($1.43/$15.00), among others. See the pricing split.
Xiaomi (小米 / MiMo)
The AI research arm of smartphone maker Xiaomi. The MiMo-V2.6 series (announced 2026-09-22) is pitched as “natively omnimodal” and ships in three variants: MiMo-V2.6-Pro (top capability), MiMo-V2.6-Flash (the intelligence/efficiency/cost balance), and MiMo-V2.6-Pro-UltraSpeed (same quality at up to 20x output speed). Its official model page confirms text, image, video and audio input, text output, and a 1M-token context window. API pricing is unchanged from V2.5: Flash $0.14/$0.28, Pro $0.435/$0.87, Pro-UltraSpeed $4.35/$8.7 (cache-miss input / output). Xiaomi claims MiMo-V2.6-Pro scores 46.32 on the Artificial Analysis Intelligence Index v4.3, “surpassing Kimi K3 and Qwen3.8 Max to become the strongest open-source model to date”. The team works “built in public”: the technical report, training environment and RL code are released, and the production RL run was live-streamed. Pricing table.
ByteDance Doubao / Baidu ERNIE / Tencent Hunyuan
ByteDance's Doubao (豆包) is a leading consumer assistant (closed). Baidu's ERNIE 5.x is omnimodal (only ERNIE 4.5 is open-source). Tencent open-sourced Hunyuan-A13B.
Model pricing & open weights (indicative)
USD per 1M tokens. DeepSeek verified on the Pricing Comparison. Others reflect published figures from official announcements/docs, with third-party reporting/evaluations attributed, as of 2026-08-29 (routers and regions may differ).
| Model | Provider | Input | Output | Open weights | Notes / source |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek | $0.15 / $0.30 (off/peak) | $0.60 / $1.20 (off/peak) | MIT | Cache miss. Released 2026-09-10. Guide |
| DeepSeek V4 Pro (live) | DeepSeek | $0.66 / $1.32 (off/peak) | $1.98 / $3.96 (off/peak) | MIT | Live. Continues after 9/14 (routing notice withdrawn). Continuation details |
| Qwen3.8-Max | Alibaba | $2.00 | $6.00 | Partial (text: custom license) | 2.4T MoE, 1M ctx. Launch pricing (MarkTechPost report) |
| Qwen3.8-Flash New | Alibaba | $0.15 | $0.47 | Open (Flash-Next) | 125B/6B active MoE, 256K→1M ctx, multimodal. Qwen4 preview. Deep dive. Qwen blog |
| Qwen3.8-27B | Alibaba | — | — | TBD | Smaller self-hostable checkpoint (license: see official model card) |
| GLM-5.3-Flash New | Zhipu (Z.ai) | $0.15 | $0.50 | MIT | 320B/18B MoE, 1M ctx. List price (intro promo ended 9/9). Deep dive. z.ai docs |
| GLM-5.2 / 5.3 | Zhipu (Z.ai) | $1.40 | $4.40 | MIT (5.2) | 5.3 weights not public as of retrieval. z.ai docs |
| Kimi K2.6 | Moonshot | $0.95 | $4.00 | Modified MIT | Reported (Miraflow). Said to match GPT-5.5 on SWE-bench Pro |
| Kimi K3 | Moonshot | $3.00 | $15.00 | Open (2.8T) | Released 2026-07-16; among the largest open-weight models. Cached input $0.30 (90% off); cache writes $3.00 (5-min TTL) or $6.00 (1-hour TTL) — all Moonshot direct API pricing; third-party hosts such as Bedrock are priced separately. 1M context, native vision. Always reasons (reasoning_effort=max), so reasoning tokens bill as output. Official pricing (verified 2026-09-25) |
| MiniMax M3 | MiniMax | $0.30 | $1.20 | Open | List $0.60/$2.40 with "permanent 50% off" (Puter/AA) |
| step-5-preview New | StepFun | $1.00 | $2.70 | — | Listed under the “Multimodal Reasoning Models” category of the official pricing table. Cache hit $0.10. StepFun docs |
| step-3.7-flash | StepFun | $0.20 | $1.15 | — | Lower-cost model in the same category. Cache hit $0.04. StepFun docs |
| MiMo-V2.6-Pro New | Xiaomi | $0.435 (cache hit $0.0036) | $0.87 | Open (built in public) | Announced 2026-09-22. Natively omnimodal (text/image/video/audio in, text out), 1M context. Pricing unchanged from V2.5; cache writes free for a limited time. Xiaomi claims 46.32 on the AA Intelligence Index v4.3. Xiaomi announcement · Model specs (verified 2026-09-27) |
| MiMo-V2.6-Flash | Xiaomi | $0.14 (cache hit $0.0028) | $0.28 | Open | The balanced model in the family: same 1M context and four input modalities as Pro. Batch API runs at 50% of real-time pricing (official docs). Xiaomi announcement (verified 2026-09-27) |
| MiMo-V2.6-Pro-UltraSpeed | Xiaomi | $4.35 (cache hit $0.036) | $8.7 | — | An API serving variant of the same Pro weights (up to 20x output speed). It costs 10x the Pro rate on both input and output, so use it selectively. Xiaomi announcement (verified 2026-09-27) |
StepFun's pricing split — the “Multimodal Reasoning Models” category
StepFun's documentation divides its price table by model category, and one of those categories is Multimodal Reasoning Models. Prices are USD per 1M tokens, with cached input billed separately (checked 2026-09-23).
| Model | Input (cache miss) | Input (cache hit) | Output | Note |
|---|---|---|---|---|
| step-5-preview | $1.00 | $0.10 | $2.70 | The higher-end model in this category; cached input is 1/10 of the standard input rate |
| step-3.7-flash | $0.20 | $0.04 | $1.15 | The lower-cost model in this category |
DeepSeek in depth
We maintain a hands-on guide for DeepSeek.
- DeepSeek V4.1 Flash guide — pricing, benchmarks, setup, FAQ.
- DeepSeek V4 Pro: continuation details — why V4 Pro keeps running and how it compares with V4.1 Flash.
- DeepSeek pricing (peak/off-peak) — current V4.1 Flash and V4 Pro rates, plus time-window details.
- Release timeline — from R1 to V4 Pro GA.
Update log
This page is updated continuously; newest entries appear first.
- 2026-10-13StepFun's official docs changed the cache-hit input price for step-5-preview from $0.05 to $0.10 (cache-miss input $1.00 and output $2.70 unchanged; cache hit is now 1/10 of standard input). Confirmed with two live fetches and corrected the pricing table. DeepSeek (V4.1 Flash / V4 Pro), GLM, Kimi K3 and MAI were re-checked and unchanged.
- 2026-09-27Added Xiaomi MiMo-V2.6 series. Verified against Xiaomi's announcement page (mimo.xiaomi.com) and official model page (mimo.mi.com): pricing (Pro $0.435/$0.87, Flash $0.14/$0.28, Pro-UltraSpeed $4.35/$8.7; cache hits $0.0036/$0.0028/$0.036), input modalities (text, image, video, audio), text output, 1M context, and the 2026-09-22 announcement date. Pricing is unchanged from V2.5 as stated officially.
- 2026-09-23Added StepFun (阶跃星辰). Sourced step-5-preview ($1.00/$2.70) and step-3.7-flash ($0.20/$1.15) from the “Multimodal Reasoning Models” category of the official pricing table (platform.stepfun.ai) and added them to the comparison table. DeepSeek pricing re-verified (unchanged).
- 2026-09-09GLM 5.3 Flash's 50% intro promo ($0.075/$0.25) ended; list price ($0.15/$0.50) now applies. Deep-dive, pricing and calculator updated.
- 2026-09-01Added Qwen3.8-Flash (Alibaba, released 2026-08-26): $0.15/$0.47 (international), 125B/6B active MoE, 256K→1M context, Qwen4 architecture preview — verified against qwen.ai/blog. New deep-dive page.
- 2026-08-31Added GLM 5.3 Flash (Z.ai, released 2026-08-26): list $0.15/$0.50 (the 50% intro promo $0.075/$0.25 ended 9/9), 320B/18B MoE, 1M context, MIT — verified against docs.z.ai and the Z.ai blog. New deep-dive page.
- 2026-09-25Added Kimi K3 pricing, verified on the official pricing page ($3.00 input / $15.00 output / $0.30 cached input, 1M context, native vision), plus its Amazon Bedrock availability (AWS blog, 2026-09-18). Also added Kimi K3 data handling via Bedrock (ZOA/ZDR, not shared with model providers, not used for training — flagged as Bedrock-wide policy, not Kimi-specific).
- 2026-08-29Page created. Added an overview of DeepSeek, Qwen, GLM, Kimi, MiniMax and other Chinese models, including DeepSeek peak/off-peak pricing and Qwen3.8-Max, GLM-5.x, Kimi K3, MiniMax M3 highlights, verified against primary sources.
⚠️ Disclaimer
- Pricing and specs prioritize official pages/announcements (with third-party reporting attributed) as of 2026-08-29 and may change without notice. Always check official pages before contracting.
- Benchmark figures exist in both self-reported and independent forms and may diverge. This page lists only what could be verified.
- This site is informational only. We are not affiliated with or endorsed by any provider. All trademarks belong to their owners.