Chinese LLMs: MiMo, Qwen, DeepSeek, GLM, Kimi

Two years ago, a Chinese model was mostly a small Qwen you ran locally out of curiosity. In October 2026, Xiaomi’s MiMo, Moonshot’s Kimi or Z.ai’s GLM code almost as well as the leading models from Anthropic and OpenAI, and often cost much less. Many of them plug straight into Claude Code, OpenCode or Codex, with a monthly plan designed for exactly that.

This guide covers it all: why they are worth a look, the three ways to use them, the exact configuration for each tool, and the awkward question, the one about your data.

Why they are worth a look

Three reasons, with numbers (rankings from Quelle IA, in French, 28 September 2026 edition).

• They are close behind the best. In the overall ranking, the top Chinese model, Xiaomi’s MiMo-V2.6-Pro, is 13th out of 246 with a score of 90.5, followed by Kimi K3 (14th), Qwen 3.8 Max (21st), GLM-5.3 (24th) and DeepSeek V4 Pro 0813 (27th). At the top, Claude Opus 5.5 scores 99.9. In coding, GLM-5.3 ranks 9th with 91.5 (Code ranking).
• They cost much less. MiMo-V2.6-Pro costs €0.44 per million tokens read and €0.87 per million written, against €4 and €20 for Claude Opus 5.5. In the value-for-money ranking, four of the top five are Chinese: three DeepSeek models and MiMo-V2.6-Pro.
• Their weights are open. The top sixteen rows of the open-weights ranking are Chinese. Open means downloadable: any host can serve them, and the smaller ones run at home. On a machine with between 22 and 189 GB of usable memory, the best-rated model that fits is Alibaba’s Qwen 3.8 27B (82.4).

Three ways to use them

1. The vendor’s coding plan

Most Chinese vendors sell a monthly subscription reserved for coding, to plug into your agent. You pay a fixed amount, the vendor counts your usage its own way, and you are cut off (or slowed down) at the cap.

Public prices in dollars, before tax, recorded on 30 September 2026 by Quelle IA from each offer’s page. One vendor’s credits cannot be compared with another’s: each has its own scale. The full table, with yearly discounts and exact models, is on the Quelle IA Tools page (in French).

2. Pay as you go: the vendor’s API or an aggregator

Every vendor also sells its API per million tokens, with no subscription: you top up a balance and pay for what you use. That is the normal route for a script, a bot or an application, and the only one at DeepSeek, which sells no coding plan. Each model’s prices are in the Quelle IA comparison tool (in French).

If you want to try several models without opening five accounts, go through an aggregator (we introduce them in Installing the agent):

• OpenRouter: one key for hundreds of models, including every model in this guide, paid as you go. Perfect in OpenCode. In Claude Code, OpenRouter warns that it is only guaranteed to work with Anthropic’s models.
• OpenCode Zen: the OpenCode team’s pay-as-you-go gateway, with models tested inside an agent. Its page says it plainly: “All our models are hosted in the US”.
• OpenCode Go: the same team’s plan, $10 a month ($40 for Go Plus), with GLM-5.3, Kimi K3, MiMo-V2.6-Pro, MiniMax M3, Qwen 3.8 and DeepSeek V4.1 Flash, among others. Each model has its own monthly cap in dollars of usage (recorded on 30 September 2026). A good way to try them all for the price of a lunch.

3. Locally on the mini-PC

This is the route this site prefers: open models run at home with Ollama, for free, without sending anything to anyone. The big Chinese models (Kimi K3, GLM-5.3, MiMo-V2.6-Pro) need several hundred gigabytes of memory: out of reach for a mini-PC. Their smaller cousins fit very well.

# 16 GB of memory: Qwen3.5-9B, a 6.6 GB download
ollama pull qwen3.5:9b

# 32 GB of unified memory or a 24 GB graphics card: Qwen 3.8 27B, 18 GB
# the best-rated model that fits, up to 189 GB of usable memory
ollama pull qwen3.8:27b

# 64 GB and up: Qwen3.5-35B-A3B, 24 GB, faster because it only activates 3 billion parameters at a time
ollama pull qwen3.5:35b-a3b

Even on a 128 GB Ryzen AI Max+ 395, Qwen 3.8 27B remains the best-rated model that fits (that machine’s page on Quelle IA, in French). A bigger model is not necessarily a better one: choose with Choosing and sizing your model and Quelle IA, machine by machine.

Plugging them into your tools

Most of these vendors expose two doors: one compatible with Anthropic’s API, which Claude Code knows how to use, and one compatible with OpenAI’s, for OpenCode and Codex. You just tell the tool which address to call and with which key.

# a settings file just for GLM, next to your normal configuration
cat > ~/.claude/glm.json <<'EOF'
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "your_z_ai_key",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash[1m]",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}
EOF
chmod 600 ~/.claude/glm.json        # the key is inside: readable by you only

# a Claude Code session on GLM; plain "claude" stays on Anthropic
claude --settings ~/.claude/glm.json

opencode
# inside OpenCode:
/connect      # search for "Z.AI Coding Plan", for instance, then paste the key
/models       # pick GLM-5.3

curl -fsSL https://mimo.xiaomi.com/install | bash   # installs the mimo command
mimo auth login                                     # pick MiMo, sign in through the browser
cd ~/projects/my-project && mimo                    # then /models to switch models

# ~/.codex/config.toml: declares the provider without making it the default
[model_providers.kimi]
name = "Kimi"
base_url = "https://api.kimi.ai/coding/v1"   # outside China; api.kimi.com in China
env_key = "KIMI_API_KEY"                     # Codex reads the key from this variable
wire_api = "responses"

# ~/.codex/kimi.config.toml: the "kimi" profile
model_provider = "kimi"
model = "k3-256k"
model_catalog_json = "~/.codex/models.json"   # model sheet supplied by Kimi

export KIMI_API_KEY="your_kimi_key"   # put it in ~/.bashrc to keep it
codex --profile kimi                  # plain "codex" stays on OpenAI

Your data: where it goes, what is done with it

An open model and an online service are not the same thing. The weights of GLM or Qwen can be downloaded and run anywhere. When you go through the vendor’s API or plan, your code goes to them, and their terms apply. Here is what those terms say, read on 1 October 2026 on the official pages:

For these international offers, the contract is signed with a Singapore company (Alibaba goes through its Dutch subsidiary for a European billing address), except DeepSeek, which contracts from Hangzhou under Chinese law.

None of this is unique to China: with any online provider, American ones included, your code leaves the machine and depends on terms you do not control. What weighs here: Chinese parent companies, terms that sometimes contradict each other from one page to the next, and little leverage over them from Europe. The right attitude fits in one rule: only send these services what you would publish without regret.