Chinese LLMs Compared: Qwen, DeepSeek, Kimi, GLM, MiniMax and ERNIE

“Every list I read says a different Chinese model wins. Which one do I actually pick?” In September 2026 the choice usually comes down to six families: Qwen, DeepSeek, Kimi, GLM, MiniMax, and ERNIE. No family wins every job. The top row moves with your workload, your hardware, your context needs, and whether you need downloadable weights. DeepSeek’s V4 line, GLM-5.3 and Kimi K3 carry open weights under their own licenses; Qwen3.8-27B and ERNIE 4.5 are Apache 2.0. A “Chinese model” label means the lab is Chinese, not that the model writes Chinese well.

Which Chinese-developed LLM families lead in 2026?

No single family leads, and that is the part worth keeping. Turing Post’s 2026 family review is an outside comparison, not MixRoute’s ranking, and it assigns current strengths this way: DeepSeek V4 for long-context reasoning and agents, Qwen3.8 for coding and multimodal work, Kimi K3 for long-horizon multi-agent tasks, and GLM-5.3 for coding and tool use. The review also states the condition that matters more than any row in it: the best choice depends on workload, hardware, context requirements, and whether you need open weights. So the useful question is not which family sits on top. It is which of those four constraints your project actually hits first.

Six families behind one “Chinese model” label

A ranking filtered by “Chinese” tells you where the lab sits, not how well the model writes the language. That gap explains a lot of the apparent disagreement between leaderboards. These are the families behind the filter:

• Qwen from Alibaba, which pairs a large hosted MoE flagship with open-weight releases at several sizes.
• DeepSeek, whose current V4 line carries MIT-licensed weights for long-context reasoning, coding, and agent work.
• Kimi from Moonshot AI, whose K3 model is open weight under the custom Kimi K3 License and built for long-horizon agent tasks.
• GLM from Z.ai, the Tsinghua-linked lab whose GLM-5.3 is a post-training upgrade of the GLM-5.2 base for software engineering, terminal work, and long-running agents.
• MiniMax, whose M3 model competes as a lower-cost multimodal worker.
• ERNIE from Baidu, which is no longer only a hosted family: Baidu released ERNIE 4.5 weights and development toolkits under Apache 2.0 while keeping a separate hosted ERNIE API.

A leader that changes with the filter

Watch the ordering move when the filter moves. The BenchLM snapshot dated September 18, 2026 lists Kimi K3 at 74.43 and Qwen3.8 Max at 73.17 under BenchAlign v5.2. Both are composite scores, not task pass rates. A composite number does not tell you whether the checkpoint you can download is the one that was scored, so the release and the deployment option behind a row decide whether it belongs on your shortlist.

How do Qwen, DeepSeek, Kimi K3, GLM, MiniMax and ERNIE compare?

Each family ships several releases, and the release you can actually call or download is the row that counts. The table below lists representative 2026 releases with the figures each lab publishes for that release. The weight column separates hosted products from downloadable checkpoints, because a family name does not tell you which release you can run or which license covers it.

What the weight and license column does not tell you

Open weight means the parameters are downloadable under a model-specific license. That is not the same as an OSI-approved license, unrestricted commercial use, or inexpensive deployment, so read each release’s terms on their own.

• Hosted Qwen3.8 Max accepts images and video. The separate downloadable Max-scale Qwen checkpoint is described as text-only with mandatory thinking. Features on the hosted product do not carry onto the downloadable artifact.
• GLM-5.2 is MIT-licensed and GLM-5.3 shares its base, but that does not make GLM-5.3 MIT. The GLM-5.3 License is its own document.
• ERNIE 4.5 shipped under Apache 2.0. That release says nothing about every hosted or future ERNIE product.

What do the benchmark scores behind these comparisons actually measure?

A score answers the question its method was built for, and nothing wider. Chinese-model comparisons mix three kinds of numbers: a public evidence-aware snapshot, a broad composite index, and vendor-reported results. Each method sets a ceiling on the claim a score can support.

BenchLM’s Supported and Estimated labels

BenchLM marks rows Supported or Estimated so you can see how much evidence sits under each one. Its overall score combines capabilities, while task-specific columns answer narrower questions. Read the method, the source dates, and the uncertainty range next to the score. A small point difference on its own is not a reason to move your production model.

The Artificial Analysis Intelligence Index is broad, not coding-only

You will also see the Artificial Analysis Intelligence Index quoted in Chinese-model posts, and it is easy to read as a coding ranking. It is not one. The index folds coding, terminal, reasoning, knowledge, and mathematics evaluations into a single composite, which makes it useful for building a shortlist and weak as the last word on a deployment. Coding decisions need coding-specific evidence gathered under your own agent setup.

Vendor-reported results come from the lab that published them

Z.ai publishes the most specific current GLM coding results, and those figures are vendor-reported. The company reports that GLM-5.3 improved from 4.6 to 28.3 on Terminal-Bench 3.0 and from 46.2 to 66.9 on DeepSWE v1.1. On Z.ai’s private Code Bench, GLM-5.3 reached 34.5% at max effort while using about 75,000 output tokens per task, and GLM-5.2 scored 23.4% while using roughly 96,000 tokens. Those numbers describe terminal work and software-engineering tasks under Z.ai’s own evaluation conditions, and they are not independent measurements. Keep that scope attached when you repeat them.

Which Chinese LLM should you use for Chinese-language writing and conversation?

None of the rows above measure Chinese output. A leaderboard row tells you where a model was built and how it scored on broad categories, not how a Simplified Chinese support reply or a Traditional Chinese product page reads. If your product ships copy that customers read in Chinese, the result you can trust comes from a prompt set that mirrors that work, scored blind, with a native speaker deciding the close calls. Published scores can inform the shortlist. They cannot settle it.

Start with families that advertise Chinese or multilingual strength

Three families say something about Chinese or multilingual work in their own materials. Alibaba’s Qwen3-generation materials advertise official support for more than 100 languages and dialects, and the Qwen3.8 line carries that positioning into coding and agent work. ERNIE is Baidu’s Chinese-market line, and ERNIE 4.5 now gives you an Apache 2.0 checkpoint to test instead of only a hosted API. GLM began as bilingual English-Chinese ChatGLM models before scaling into larger MoE architectures. Qwen, ERNIE, and GLM are reasonable places to start a Chinese-language trial, and none of them is a proven winner for your register and your audience.

Run a blind pass with your own production prompts

Build a small prompt set that mirrors what your product does: conversational replies, customer support responses, product copy, technical documentation, and any long-form Chinese writing your team produces. If your audience spans both writing systems, cover both, because Traditional and Simplified Chinese carry separate register and terminology conventions. Pull domain vocabulary from your own work, apply the same system message and tone constraints you use in production, and score the outputs with the model names hidden so they cannot bias the grade. Where tone or cultural fit decides the outcome, a native speaker reviews the drafts. No published ranking score settles Chinese-language quality, so the blind pass runs before you commit to a model.

Three prompts you can reuse in a Chinese-language trial

Treat these as starting inputs and swap the fictional details for your own. Keep the model names hidden from the reviewer and mark each output for factual accuracy, regional wording, format compliance, and the edits needed before use.

A run only compares to the next one when the model ID, prompt, settings, latency, reported usage, and reviewer edits travel with it. These prompts define a trial. They do not report model results.

If DeepSeek is still on your shortlist, see how DeepSeek and ChatGPT differ on everyday language and reasoning tasks, then put both through the same blind pass before you commit.

Which Chinese LLM should you pick for coding and long-running agents?

The coding shortlist is a set of trial candidates grouped by workload and by license or hosting constraint, not a ranking. Compare accepted results under the same agent setup, because a coding leaderboard does not tell you which model finishes patches inside your tool harness.

The coding shortlist, by workload and constraint

• Text-only repository and terminal loops: GLM-5.3 is the candidate to test first for long-running software engineering and terminal work. Z.ai’s vendor-reported results support that role, and official GLM-5.3 weights are now downloadable under the GLM-5.3 License. If your team also weighs the earlier MIT-licensed GLM-5.2 release, compare GLM-5.2 with Claude Opus 4.8 and Fable 5 before setting a default.
• MIT-licensed self-hosting for backend work: DeepSeek V4 Pro 0813 is the permissive-weight candidate, with a 1M context and thinking and non-thinking modes on the API. Check the exact output setting for the route you use; a large maximum output recommended for local high-effort reasoning is not automatically the hosted API cap.
• Frontend, screenshot, and video work: Qwen3.8 Max accepts text, images, and video and returns text, which fits rebuilding an interface from a screenshot or debugging a rendered frontend against a reference.
• Difficult multimodal coding with open weights: Kimi K3 is the high-capability candidate when visual evidence or very hard reasoning justifies its cost and infrastructure weight. Confirm the Kimi K3 License and the current official deployment guide before choosing it, because a 2.8T-parameter model needs substantial multi-GPU infrastructure.
• High-volume, lower-cost subtasks: MiniMax M3 is positioned as a budget multimodal worker for triage and first-pass review. Verify current route rates and license terms instead of assuming the whole family is always the cheapest option.

The model and the agent are a package

A coding benchmark winner does not automatically become the best coding agent. The agent decides which repository files enter the context, how tool output is presented, which edit formats are allowed, when tests run, what happens after a failed command, and how much context survives between turns. Run the same model inside two different agent systems and the completion rate can move a long way, so pick the model and the agent together. If your team already runs an agent harness, that harness is the thing to hold constant while you swap models. A one-million-token context window is also not an instruction to paste an entire repository into every request. Generated files, old logs, and unrelated documentation pull attention away from the execution path that matters.

Should you self-host an open-weight Chinese LLM or call an API?

Settle the hosting and licensing question before you compare capability numbers, because a model you cannot run or cannot reach does not solve your problem. The open-weight Chinese families do not share one license or one hardware profile, and the API routes that expose them are not identical to the upstream model cards.

Where the open-weight licenses split

• DeepSeek: the current V4 line is MIT-licensed, which gives you a permissive self-hosting option when MIT is the license you want. Apache-2.0 releases are permissive in the same way.
• Qwen: Alibaba licenses per release. Qwen3.8-27B is Apache 2.0, while hosted Qwen3.8 Max and the downloadable Max-scale Qwen checkpoint stay separate artifacts with separate terms.
• GLM: Z.ai offers official GLM-5.3 weights under the GLM-5.3 License. GLM-5.3 shares its base with GLM-5.2, so the older version number does not make the older model smaller or cheaper.
• Kimi: K3 is downloadable under the custom Kimi K3 License.
• MiniMax: M3 sits under the MiniMax Community License, which needs a read before redistribution or a model-as-a-service offering.
• ERNIE: Baidu released ERNIE 4.5 weights and toolkits under Apache 2.0, while the hosted ERNIE API remains a separate commercial route.

What self-hosting actually costs

Downloadable weights are not the same as cheap deployment. Hardware requirements vary enormously across this set: some Qwen open releases and other smaller models can run on a single workstation, while trillion-parameter MoE checkpoints such as Kimi K3 at 2.8T require substantial multi-GPU infrastructure for practical inference. If you want a Max-scale Qwen on your own hardware, use the separate downloadable Max-scale checkpoint, because the hosted route is not a self-hosting entry. The two also differ in modality: hosted Qwen3.8 Max accepts text, images, and video and returns text, while the downloadable Max-scale Qwen checkpoint is text-only with mandatory thinking.

An open checkpoint also says nothing about the cost of running it well, and an open-weight row on a leaderboard does not mean the model is easy to host. For a team without that infrastructure, a hosted API can be the cheaper operational choice even when the weights are free to download. Compare current per-token rates for the exact release and route you plan to use, and remember that reasoning tokens, cache state, retries, and output length all affect the final bill.

API routes can be narrower than the model

A route a host exposes is not the model’s full specification. One portal route may serve GLM-5.3 as text-to-text while another host adds multimodal input or a different context ceiling. The same gap runs inside a single family: the hosted Qwen3.8 Max route accepts images and video, while the downloadable Max-scale Qwen checkpoint is text-only with mandatory thinking. Read the route documentation before you promise a feature to your team, and treat the model card as the starting point rather than the contract.

Once your model list and deployment path are set, the remaining decision is where to buy API access. MixRoute’s model catalog documents exposed model IDs, API modes, and context specifications for the models it lists, and the catalog is a reference, not a quality ranking or a proof of Chinese-language suitability. Verify the exact model ID, modality, context limit, and API mode for the release you choose. Decide whether a MixRoute prepaid credit tier fits your model budget before you commit to a provider.