Chinese LLMs in November 2026: Qwen4-Max Reprices the Top Tier, Hy4 Flash Drops the Floor to CNY 0.50
A complete side-by-side of the current Chinese LLM lineup, priced from the vendors' own pages on November 17: Alibaba Cloud Qwen4-Max at CNY 10 / CNY 30, Tencent Hunyuan Hy4 flash at CNY 0.50 / CNY 2.00, Qwen3.8-Max at CNY 12 / CNY 36, Zhipu GLM-5.3 at CNY 8 / CNY 28, Moonshot Kimi K3.1 Pro at CNY 13 / CNY 68, ByteDance Doubao-Seed-2.1-Pro at CNY 6 / CNY 30 and Xiaomi MiMo-V2.6-Pro at CNY 3 / CNY 6. Includes the full cached-input price ladder, the token-cost effect of Alibaba's new Qwen4-Max tokenizer, and a one-page selection guide.
Two releases on November 16 moved both ends of the Chinese model price spectrum in one afternoon. Alibaba's Qwen4-Max repriced the flagship tier down to CNY 10.00 input and CNY 30.00 output with a CNY 0.50 cache hit rate, and on Terminal-Bench 4.0 it used 151K tokens per attempt against Qwen3.8-Max's 289K while the score rose from 37.9 to 48.5 - a 58 percent per-task cost drop, larger than the 16.7 percent price cut. Tencent's Hy4 flash pushed the floor to CNY 0.50 input, CNY 2.00 output and CNY 0.02 cached, with a 1M window, multimodal input and decoding above 150 tokens per second. Everything else was re-verified with zero movement. Here they all are side by side, with the cached-input ladder that shows up on an agent invoice.
The new floor: Hy4 flash at CNY 0.50
Tencent Hunyuan Hy4 flash lists CNY 0.50 input, CNY 2.00 output and CNY 0.02 cached per million tokens - the cheapest CNY input tier and the cheapest 1M multimodal option available. The cache rate is 4 percent of standard input and the KV cache footprint is one twenty-fourth of Hy3, which matters more than the headline rate for any loop that re-reads a large context. Parameters are 224B total with 12B active, decoding runs above 150 tokens per second. Through Tencent Cloud TokenHub it is at 50 percent until November 30: CNY 0.25 / CNY 1.00 / CNY 0.01. The previous CNY 0.80 floor was shared by four models, all also 1M-context multimodal: Qwen3.8-Flash, Qwen3.8-Omni-Flash, Doubao-Seed-2.1-Lite and GLM-5.3-Flash. On a 100M-token monthly batch, Hy4 flash at the promo rate costs CNY 25 against CNY 80 for the CNY 0.80 tier.
The top tier repriced: Qwen4-Max at CNY 10
Qwen4-Max lists CNY 10.00 input and CNY 30.00 output in China, against CNY 12.00 and CNY 36.00 for Qwen3.8-Max - a 16.7 percent cut on both sides - with implicit cache hits at CNY 0.50. Explicit cache creation is CNY 12.50 for five minutes and CNY 20.00 for one hour. The international tier runs USD 1.50 / USD 4.50 with USD 0.075 cached, and Batch calls are half price. The architecture is a 2.4T-parameter MoE using Qwen Sparse Attention with native vision, a 1M window and 128K maximum output. What matters more than the price cut is token efficiency: on Terminal-Bench 4.0 at the same Xhigh tier, average tokens per attempt fell from 289K to 151K while the score rose from 37.9 to 48.5. Multiply those together and a flagship workload costs roughly 58 percent less per task.
The full ladder and the cached-input ladder
Normalizing to USD at roughly CNY 7.16 per dollar gives a clean ladder. Bottom rung: Hy4 flash at CNY 0.50, then the CNY 0.80 group of Qwen3.8-Flash, Qwen3.8-Omni-Flash, Doubao-Seed-2.1-Lite and GLM-5.3-Flash, all 1M-context multimodal. The CNY 1 tier holds Spark X2.6 at CNY 0.90, MiMo-V2.6-Flash at CNY 1.00 and Tencent Hy3 at CNY 1.00 but only 256K, plus DeepSeek V4.1 Flash at USD 0.15 off-peak. Mid tier: GLM-5.3-FlashX CNY 2.00, MiMo-V2.6-Pro CNY 3.00, Doubao-Seed-2.1-Turbo CNY 3.00 on 256K. Flagship tier runs from Doubao-Seed-2.1-Pro at CNY 6.00 up to Kimi K3.1 Pro highspeed at CNY 26.00 / CNY 136.00, with Qwen3.8-Max demoted to CNY 12.00 / CNY 36.00. On cached input, which is what an agent invoice actually shows: Hy4 flash CNY 0.02, Qwen3.8-Flash and Omni-Flash CNY 0.10, GLM-5.3-Flash CNY 0.23, Qwen4-Max CNY 0.50 against Qwen3.8-Max's CNY 1.50, Doubao-Seed-2.1 Pro CNY 1.20, and Kimi K3.1 Pro CNY 1.30 with writes billed separately by TTL at CNY 13.00 and CNY 26.00. A Kimi hit is not re-billed for the write.
Which one to pick, in one page
Cheapest 1M-context input: Hy4 flash at CNY 0.50, or CNY 0.25 through TokenHub until November 30. Cheapest flagship-quality Chinese model: Qwen4-Max at CNY 10.00, because a Qwen3.8-Max pipeline costs roughly 2.4 times more per task. Cheapest long-context agent loop: Hy4 flash at CNY 0.02 cached. Cheapest batch processing: DeepSeek V4.1 Flash off-peak at USD 0.15 / USD 0.60. Highest published output ceiling in China: Kimi K3.1 Pro highspeed at CNY 136.00, with the standard Kimi K3.1 Pro at CNY 68.00 and open-source-friendly DeepSeek V4 Pro at USD 0.66 / USD 1.98 off-peak. Cheapest native multimodal: Qwen3.8-Omni-Flash at CNY 0.80. And if you need one model that survives the next two quarters, Qwen4-Max with a 1M window, 128K output and native vision is the answer - but it is not the cheapest way to serve that volume, which remains Hy4 flash's job.
Frequently Asked Questions
What is the cheapest Chinese LLM API right now?
Tencent Hunyuan Hy4 flash at CNY 0.50 input, CNY 2.00 output and CNY 0.02 cached input per million tokens. Through Tencent Cloud TokenHub it runs at 50 percent until November 30, 2026, which makes the effective rate CNY 0.25 input, CNY 1.00 output and CNY 0.01 cached. The previous floor of CNY 0.80 was shared by Qwen3.8-Flash, Qwen3.8-Omni-Flash, Doubao-Seed-2.1-Lite and GLM-5.3-Flash. Note that Alibaba's free-tier decision model lists CNY 0 on both directions, but that is a gift quota rather than a pay-per-use tier and is not comparable.
Is Qwen4-Max cheaper than Qwen3.8-Max in practice?
Yes, and by more than the price table shows. Qwen4-Max lists CNY 10.00 input against CNY 12.00, and CNY 30.00 output against CNY 36.00, a 16.7 percent cut on both sides, with cached input dropping from CNY 1.50 to CNY 0.50. On top of that, on Terminal-Bench 4.0 at the same Xhigh tier, average tokens per attempt fell from 289K to 151K while the score rose from 37.9 to 48.5. Combined, the average per-task cost falls by about 58 percent, so an existing Qwen3.8-Max pipeline costs roughly 2.4 times more per task than the same pipeline on Qwen4-Max.
Does every current Chinese flagship have a 1M context window?
Almost, with two exceptions worth knowing. DeepSeek V4 Pro, Zhipu GLM-5.3, Alibaba Cloud Qwen4-Max and Qwen3.8-Max, Moonshot Kimi K3 and Kimi K3.1 Pro, ByteDance Doubao-Seed-2.1-Pro, Xiaomi MiMo-V2.6-Pro and Tencent Hy4 flash all sit at 1,048,576 tokens. The exceptions are Tencent Hy3 on its API tier at 256K, Volcengine Doubao-Seed-2.1-Turbo at 256K, and Xiaomi MiMo-V2.6-Flash which is a lower tier. The Kimi K2 family, which still appears in many price lists, caps at 262,144 tokens.
How should I pick between Hy4 flash and Qwen4-Max?
They serve different tiers rather than competing directly. Hy4 flash is the cost tier: CNY 0.50 input, CNY 0.02 cached, 150-plus tokens per second decoding, a 1M window and multimodal input. Use it for bulk classification, routing, log summarising, document pre-filtering and high-turn customer support. Qwen4-Max is the capability tier: a 2.4T-parameter MoE with native vision, 1M context, 128K output, and the strongest coding and reasoning numbers in the domestic lineup. Use it for agentic coding, long-horizon refactors, research over a full corpus and Chinese business documents. A common production pattern is Hy4 flash for the cheap first pass and Qwen4-Max for the expensive second pass.
Want more AI pricing comparisons?
Check our price changelog for daily updates on AI model pricing, new releases, and deprecations.