Three Chinese labs shipped flagship models within two weeks of each other. Alibaba released Qwen4-Max on November 16, 2026, Moonshot AI put Kimi K3.1 Pro weights on Hugging Face in early November after an October API launch, and Tencent fully released Hy4 on November 4. They overlap on a 1M-token context window and agent-style coding work, but they differ a lot on openness, price and maturity. This guide compares what each vendor has actually published so you can pick one to test.
Quick comparison
| Model | Maker | Released | Weights | Reported API price (per 1M tokens) |
|---|---|---|---|---|
| Qwen4-Max | Alibaba Cloud | Nov 16, 2026 | Closed, hosted only | $1.50 input / $4.50 output |
| Kimi K3.1 Pro | Moonshot AI | API Oct 23-24, weights early Nov 2026 | Open weights, custom license | $2 input / $10 output |
| Tencent Hy4 | Tencent | Nov 4, 2026 | Preview weights Apache 2.0, full weights announced | $0.834 input / $2.501 output |
Prices change often, and the Kimi and Tencent figures come from press coverage of the launches, so confirm them on each vendor's pricing page before you budget.
Qwen4-Max: Alibaba's closed flagship
Alibaba describes Qwen4-Max as the most capable model in the Qwen family, aimed at coding, office productivity, frontend design and complex agentic tasks. It is multimodal, has a one-million-token context window, and thinking is on by default at a high reasoning effort, with low, high and xhigh settings available. The model ID in Alibaba Cloud Model Studio is qwen4-max, and it also runs inside Qwen Studio, QwenWork and Qoder.
The list price is $1.50 per million input tokens and $4.50 per million output tokens on the international (Singapore) endpoint, about 25% below the previous flagship Qwen3.8-Max. New Model Studio users in Singapore also get a one-million-token free quota that lasts 90 days.
The trade-off is openness. Qwen4-Max is a proprietary hosted model, and Alibaba has only said that smaller Qwen4 models are planned as open source. Alibaba's model pages also do not publish a parameter count or benchmark table, so you will need to evaluate it on your own tasks.
Best for
- Teams already on Alibaba Cloud or using Qwen apps
- Long-document and office workflows that need a 1M-token window at a low list price
- Developers who do not need to self-host
Kimi K3.1 Pro: open weights for coding agents
Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, a post-trained upgrade of Kimi K3 with a one-million-token context window. Moonshot positions it for long-horizon coding, agentic work and knowledge work. It launched through the Kimi apps, Kimi Code and the Kimi API on October 23-24, 2026, alongside K3.1 Pro Highspeed, a latency tier that Moonshot says is two to three times faster with the same capabilities.
Weights followed on Hugging Face in early November under a custom Kimi license. The license is not a standard open-source license, and commercial use above certain thresholds may require a separate agreement, so read it before building a product on self-hosted weights. The base K3 is a mixture-of-experts model of roughly 2.8 trillion parameters, which means self-hosting needs serious GPU infrastructure.
Reported API pricing is $2 per million input tokens and $10 per million output tokens for the standard tier, which is higher than Qwen4-Max and Tencent Hy4. Moonshot also reports a drop in claim-level hallucinations from 23.4% for K3 to 8.3% on its own internal test, which is a vendor claim rather than an independent result.
Best for
- Developers who want the option to self-host open weights
- Coding and agent workflows through Kimi Code or an OpenAI-compatible API
- Teams that need a faster tier and will pay for lower latency
Tencent Hy4: the low-price option with vision and computer use
Hy is the new name for Tencent's Hunyuan model family. Hy4 first appeared as a preview on August 28, 2026, described as a 770-billion-parameter mixture-of-experts model with 49 billion active parameters and a context window above one million tokens, released with open weights under Apache 2.0. The full release on November 4 added vision and computer-use capabilities and is available through hosted APIs and Tencent products such as WorkBuddy.
Tencent reported API pricing of $0.834 per million input tokens and $2.501 per million output tokens, unchanged from the preview, with a 50% discount on Tencent Cloud TokenHub through December 31. That makes Hy4 the cheapest of the three on paper.
At launch, Tencent said final weights and a system card would be released shortly, and a separate Hy4 Flash model aimed at lower cost and faster inference is planned. Check whether the final weights have shipped before you rely on them. Benchmark numbers Tencent published for the preview, such as 65.7 on SWE-bench Pro, are vendor-reported.
Best for
- Cost-sensitive agent and coding workloads
- Teams that want vision and computer use from a Chinese open-weight family
- Users inside Tencent's ecosystem
How to choose
- You need open weights: Kimi K3.1 Pro or Tencent Hy4, after reading each license and confirming Hy4's final weights are out.
- You want the lowest price: Tencent Hy4, with Qwen4-Max next.
- You want one hosted API with a large free quota: Qwen4-Max on Model Studio.
- You want a faster tier for latency-sensitive agents: Kimi K3.1 Pro Highspeed.
None of these vendors publishes fully independent comparisons yet, so run a small test with your own prompts and measure quality, latency and cost per finished task rather than relying on launch benchmarks. Also check regional access, data-handling terms and whether your use case is allowed under each license.
Where to read more on AI Tech Hub
You can see the full specs for each model on our Chinese AI models page, and compare them with other tools in the Chinese AI hub. We update those pages when vendors change pricing or release new weights.
Research sources
References used while preparing this article. Verify time-sensitive details at the original source.
Found this helpful?
Share it with your network
Tags
You Might Also Like
AI NewsDoubao Phone Assistant Makes App Automation a Permission Question
IT之家 reports a consumer Doubao phone assistant and a GUI cooperation protocol under which third-party apps can allow or reject AI automation. That permission boundary matters for messages, payments and private data: “AI can use apps” should
Nov 5, 2026
AI NewsJimeng Short Drama Studio Brings Script, Image and Storyboard Models Together
IT之家 reports that ByteDance’s Jimeng launched a Short Drama Studio with DreamScript for scripts, Anycook for images and DreamShot for storyboards. The reported workflow covers multiple drama themes, long-text analysis and story-structure di
Nov 5, 2026
AI NewsKling 4.0: Longer Clips Are Only One Part of the Upgrade
Kling 4.0 has reportedly moved to full rollout on klingai.com. IT之家 says it can generate videos up to 30 seconds, while 4K and 1080p 10-bit HDR output was described as coming soon. MCP/CLI access and an API for enterprise users and develope
Nov 5, 2026
