Chinese Open-Weight Models in 2026: DeepSeek, Kimi, Qwen, GLM, MiniMax, StepFun

Last tested and updated: June 2026

Claude is expensive. Six Chinese open-weight labs cover the gap — but they aren’t all “open” in the way you think.

The hook

Three terms get thrown around as if they meant the same thing. They don’t.

• Open-source — training data (or a recipe), training code, fine-tuning scripts, and weights are all public. Truly open-source models are rare; most “open” models don’t qualify.
• Open-weight — only the trained parameters are downloadable. You can run, fine-tune, and host the model yourself. The data and training recipe stay proprietary.
• Free API — a hosted endpoint at no cost, usually with rate limits. Weights may be proprietary (GPT, Gemini) or partially open (some Kimi tiers). You pay with usage limits, not dollars.

Assume “open-weight = free” and the API bill arrives. Assume “open-weight = open-source” and the training data is locked away. Most Chinese models in 2026 are open-weight with permissive commercial use. Useful, but not “free.” Not “fully open-source” either.

The mental model

Six Chinese open-weight labs matter in 2026. They are not interchangeable. Each wins a workload, ships under a license tier, and has a cost profile once real traffic hits.

The labs at a glance

Lab	Flagship	License tier	What it wins at
DeepSeek	V4 Pro / V4 Flash	MIT-style	Tool-call reasoning, agent runtimes
Kimi (Moonshot)	K2.7 / K2 Thinking	Modified MIT	UI generation, vision, transparent billing
Qwen (Alibaba)	Qwen3 / Qwen3-Max	Apache 2.0 (most)	Long context (1M), multilingual, self-host
Zhipu (GLM)	GLM-4.5 / GLM-4.5 Air	Custom (weights OK)	Enterprise tool use, structured output
MiniMax	M3 / M3-Plus	MMLU Open	Cost-efficient generalist, BYOK default
StepFun	Step-3 / Step-3 Reason	StepFun Open	Math, code reasoning, extended chain-of-thought

Apache 2.0 (Qwen) and MIT-style (DeepSeek) let you ship commercially with few strings. Modified MIT (Kimi) and Custom (GLM) require reading the model card. Check the license pill first.

How the licensing tiers stack

1. Apache 2.0 / MIT — train, fine-tune, redistribute, ship commercially. Almost no friction. (DeepSeek, Qwen)
2. Modified MIT / “weights OK” — commercial use with conditions (attribution, use-case limits, scale thresholds). Read the card. (Kimi, GLM)
3. StepFun Open / MMLU Open — newer licenses from the labs; usually permissive but less battle-tested. (MiniMax, StepFun)
4. Hosted-only, free API tier — weights not released; you call the endpoint and accept the rate limit.

Default to Apache 2.0 or MIT if you don’t yet know which license you’ll need.

Pick your tool

Match the lab to the workload, not the price.

What to use each lab for

Pick DeepSeek V4 Pro if…

• You’re wiring up an agent harness — hermes/L01 — What Is Hermes Agent? is the entry point.
• Your workload has lots of tool calls and you want the model to think inside the tool call, not after it.
• You want cached tokens to carry across projects; DeepSeek caches repeated context, which compounds savings.

Pick Kimi K2.7 if…

• Your workload is UI generation, dashboards, or vision-heavy prompts.
• You want a monthly plan with transparent request counts (Claude’s billing is opaque; Kimi is the opposite).
• You’re willing to launch /swarm or orchestrate parallel agents manually. Kimi is weak as a single agent; strong in parallel.
• You do crypto or finance work; some orderflow platforms accept Kimi API keys only (verify with your platform’s docs).

Pick Qwen if…

• You want to self-host the weights on your own hardware. Qwen ships from 0.6B up to 480B; pick the size that fits your GPU.
• You need long context — Qwen3 supports 1M tokens, the right tool for “summarise this whole codebase” or “read these 50 PDFs.”
• Your workload is bilingual or Chinese-leaning. Qwen’s English/Chinese balance is the strongest in the open-weight tier.

Pick Zhipu (GLM) if… your workload is enterprise integration — function calls, schema-validated output, strict tool-use discipline. GLM is the most “enterprise-shaped” of the six.

Pick MiniMax M3 if… you want a reliable default for daily Q&A. It ships 7B–20B distillations that run on one consumer GPU. It plays well with BYOK (Bring Your Own Key) harnesses like Hermes.

Pick StepFun if… your workload is math or code reasoning. Step-3 Reason mode is strong on olympiad-style problems.

When to skip Chinese open-weight entirely

• You need Mythos-class intelligence. Use Claude Opus or GPT 5.x — L02 — The Anthropic Family.
• You have a Claude-validated workflow you don’t want to retest. Switching costs real time.
• You have a politically sensitive prompt. If you publish, check the model’s content policy.

L04 — Cost-Efficient AI Model Tier · L05 — Frontier vs Open-Weight

Try it

The exercise

Pick one task you actually do this week. Real beats synthetic. Run it on two of the six labs above. The point isn’t to pick the winner. The point is that the labs are not interchangeable.

How to do the comparison without burning the afternoon

1. Pick a workload you’ll actually use the output of. Don’t pick a synthetic benchmark — L01 explained why those mislead.
2. Pick two labs. Default pair: DeepSeek V4 Pro + Kimi K2.7. That pair covers tool-call reasoning and UI/vision, where beginners feel the difference fastest.
3. Write the prompt once. Same prompt, same input data, same time of day.
4. Run it twice. Note latency, output quality, cost, and whether the model asked a clarifying question or just guessed.
5. Pick your default. After two runs you’ll know which model fits this workload. Don’t generalise from one task — but don’t dismiss a lab from one task either.

A note for harness users

If you’re already running Hermes (hermes/L01 and hermes/L02), the workflow is even simpler. If you run Hermes, swap the model name in your config, restart, and rerun. You only pay tokens.

Check your understanding

Quiz: see quiz.json (6 questions, valid JSON).

What’s next

L04 — Qwen, MiniMax, GLM · L05 — Frontier vs Open-Weight · hermes/L01 — What Is Hermes Agent?