Skip to main content
VluxTech
Blog · AI

DeepSeek, Qwen, Kimi or GLM? Chinese AI Models Compared

Chinese AI models now cost a fraction of the big American ones, and some of them are free to download. An AI agent is a program that does tasks for you. If you run one, like Hermes Agent or OpenClaw, any of these models can be its "brain".

But which one? This post compares the four leading Chinese models on price, privacy and what each is good at. At the end, we say which one fits which job, and which one to put inside Hermes or OpenClaw.

The four models

We picked the newest flagship from four Chinese labs. Each sells its model through its own API, the connection software uses to talk to a model. Each also lets anyone download it.

  • DeepSeek V4.1 Flash (DeepSeek), released 10 September 2026 (announcement). Free to download under the MIT licence.
  • Qwen3.8 (Alibaba). Comes as Qwen3.8-Max, the big one, and Qwen3.8-Flash, the cheap one. A smaller version, Qwen3.8-27B, is free to download under the Apache 2.0 licence (model page).
  • Kimi K3 (Moonshot AI), the largest of the four. Its weights are public under Moonshot's own Kimi K3 licence (model page).
  • GLM-5.3 (Z.ai). Its weights are public under Z.ai's own licence, and the smaller GLM-5.3-Flash uses the MIT licence (model page).

We left out MiniMax and StepFun to keep this short. Both are worth watching.

What they cost

These are the official prices per 1 million tokens, input then output. A token is a small piece of a word.

  • DeepSeek V4.1 Flash: US$0.30 in, US$1.20 out. That's half price outside DeepSeek's busy hours (pricing).
  • Qwen3.8-Flash: US$0.15 in, US$0.47 out. Qwen3.8-Max: US$2 in, US$6 out. Both come with 1 million free tokens for 90 days (pricing, Singapore region).
  • GLM-5.3: US$1.40 in, US$4.40 out. GLM-5.3-Flash costs US$0.15 in, US$0.50 out (pricing).
  • Kimi K3: US$3 in, US$15 out (pricing).

All four charge much less when the same text is sent again, which agents do all the time. DeepSeek's busy hours are 08:00 to 11:00 and 13:00 to 17:00 WIB on weekdays. Outside those, its price halves.

In Rupiah, 1 million tokens in plus 1 million out costs about:

  • Qwen3.8-Flash: Rp11.000
  • DeepSeek V4.1 Flash, full price: Rp27.000
  • GLM-5.3: Rp104.000
  • Qwen3.8-Max: Rp144.000
  • Kimi K3: Rp324.000

That's at about Rp18.000 per US dollar (Bank Indonesia rate).

Where your data goes

This matters if customer names, orders or chats pass through the model.

  • DeepSeek: its privacy policy says it collects, processes and stores personal data in China (privacy policy).
  • Qwen: on Alibaba Cloud you choose the region. The prices above are for the Singapore region.
  • Kimi: Moonshot's platform stores data on servers in Singapore. Its policy says user content is used to train its models. It doesn't say whether API use is excluded (privacy policy).
  • GLM: Z.ai says customer data is generally processed in Singapore. It also says it doesn't store what you send through the API (privacy policy).

Because all four can be downloaded, other companies also host them. That gives you a way to use the same model without sending data to China.

What each is good at

All four say their model is strong at coding and long, multi-step tasks. Those scores come from the companies themselves, not from independent tests, so treat them as claims.

  • DeepSeek V4.1 Flash is the cheap all-rounder. It also understands images. DeepSeek says Flash is now its main model: until V4.1-Pro launches, requests to its Pro model are answered by Flash.
  • Qwen3.8 gives you two levels from one company: Flash for cheap daily work, Max for hard jobs. Both read up to 1 million tokens at once.
  • Kimi K3 is built for long jobs with little supervision, like research or long coding sessions. It also reads up to 1 million tokens at once, and it's the priciest by far.
  • GLM-5.3 is aimed at coding and agent work, and costs less than Qwen3.8-Max or Kimi K3.

When to choose what

  • Choose DeepSeek V4.1 Flash for drafts, summaries and internal notes, where cost matters most and no customer data is involved. Run big jobs outside busy hours for half price.
  • Choose Qwen3.8-Flash for cheap daily work when you want your data processed in Singapore. Move up to Qwen3.8-Max for harder jobs on the same account.
  • Choose GLM-5.3 for an agent that works with tools, files and messages, at a mid-range price.
  • Choose Kimi K3 for long, hard jobs where getting it right matters more than the bill. Don't use it for customer data until you've checked its training terms.

Which model to use with Hermes or OpenClaw

Hermes Agent and OpenClaw are free AI agents you run yourself, and you choose the model they use. We compared them in this post. Both connect directly to all four of these companies.

OpenClaw's own security guide is clear on this. Use "the latest-generation, best-tier model for any bot that can run tools or touch files/networks", it says.

It also warns that smaller, cheaper models "are more susceptible to tool misuse and instruction hijacking" (OpenClaw prompt injection guide). Hermes, for its part, needs a model that can read at least 64,000 tokens at once, and all four can (Hermes providers).

So our pick:

  • Main model for an agent with tools: GLM-5.3. It's a top-tier model at a mid-range price, aimed at agent work, and Z.ai processes data in Singapore. If the budget allows and the jobs are long, Kimi K3 is the stronger step up. Qwen3.8-Max sits between the two.
  • Background jobs: a Flash model. Hermes can use a second, cheaper model for side jobs like shrinking long chat history. DeepSeek V4.1 Flash, Qwen3.8-Flash or GLM-5.3-Flash keep that cost low. Hermes says this helper model must read at least as much text at once as your main one.
  • Chat only, nothing else: a Flash model is fine. OpenClaw's guide says smaller models are usually fine for chat-only assistants with trusted input and no tools.

To switch, put the company's API key in the agent's settings and choose the model.

The names Hermes uses are zai for GLM, kimi-coding for Kimi, deepseek for DeepSeek and alibaba for Qwen. In OpenClaw they're listed as Z.AI, Moonshot AI, DeepSeek and Alibaba Model Studio. Check the exact model name in each company's dashboard, because they change with every update.

What this means for you

  • Trying an agent cheaply? Start with Qwen3.8-Flash and its free tokens, or DeepSeek outside busy hours.
  • Giving the agent tools, files or messages? Use GLM-5.3, Qwen3.8-Max or Kimi K3, not a Flash model.
  • Customer data involved? Pick a model processed in Singapore, check the training terms, and keep a person approving what's sent.

Want help setting up an agent with the right model? See our business automation service or book a free consultation.