Top 5 Large Language Models in China

What "open" means in this list

Every model here publishes downloadable weights, and that is roughly where the similarity ends. DeepSeek's V4 weights and GLM-5 are MIT. Qwen's 27B is Apache 2.0. Kimi and MiniMax use their own licences, which allow commercial deployment but add conditions you should read before shipping. None of the five publishes its training data, and most release a technical report rather than a full training pipeline. "Open weights" is the accurate phrase; "open source" is the marketing one.

The order below follows the jobs readers bring to us most often, not a benchmark average. A model that wins one clear task is more useful than one that is second everywhere.

1. DeepSeek — frontier reasoning at the lowest price

DeepSeek moved to a 1M-token default context in April 2026 and now ships two tiers: V4-Pro at 1.6 trillion total parameters with 49 billion active, and the much smaller V4-Flash. The Flash tier is where most people actually are, and it has been refreshed twice since — V4.1-Flash in September, then again on 17 October. Weights are MIT, and the API accepts both OpenAI and Anthropic request formats, which is why DeepSeek turns up inside third-party coding agents without any glue code.

• Best for: reasoning-heavy work, agentic coding, and anyone paying per token.
• Watch out: V4-Pro is not a self-hosting proposition for a normal team. Use Flash locally, or the API.

2. Qwen — the family you can actually deploy

Qwen's advantage is not one model but a ladder of them. Qwen3.8-Max is the 2.4-trillion-parameter flagship, and Alibaba has released its weights. But Qwen3.8-27B is the model most teams end up running: Apache 2.0, 262K native context extendable to 1M, and small enough to run quantised on a workstation. Alibaba counts more than 460 open models and over three billion downloads, which is why Qwen has the deepest fine-tuning and tooling ecosystem of the five.

• Best for: private deployment, fine-tuning, multilingual work, and teams that need a size for every budget.
• Watch out: the naming is a maze — Plus, Max, Flash, Coder, A3B. Choose by parameter count and licence, not by label.

3. GLM — coding agents that produce files

Zhipu's GLM-5 arrived in February 2026 at 744 billion total parameters with 40 billion active, a 200K context and an MIT licence, and the line has since moved on to GLM-5.3. It is built for what Zhipu calls agentic engineering: a model that holds a plan across a long coding session instead of answering one prompt. The difference shows up in the output — GLM-5 writes .docx, .xlsx and .pdf files end to end — and in where it runs, with support inside Claude Code, OpenClaw and most of the open coding agents.

• Best for: coding agents, long multi-step engineering, and document generation where the deliverable is a file.
• Watch out: 200K is the smallest context on this list, so very large repositories or document sets need chunking.

4. Kimi — the largest open model

Moonshot released Kimi K3 on 27 July 2026 and opened it completely: 2.8 trillion parameters, a 1M-token context, native vision, and the technical report and training infrastructure published alongside the weights. Moonshot calls it the largest open-weight model released by anyone, and it is priced like a frontier model rather than a budget one. A smaller sibling, K2.6, arrived in October at 1 trillion parameters under a modified MIT licence, with an agent-swarm mode that coordinates up to a thousand sub-agents in parallel.

• Best for: the hardest long-horizon engineering work, very large codebases, and visual-to-code tasks.
• Watch out: 2.8T parameters is a cluster, not a workstation, and the K3 licence is Moonshot's own — read it before commercial use.

5. MiniMax — long context and video input at low cost

MiniMax M3, released in June 2026, is a 428-billion-parameter model with 23 billion active, a 1M-token context and native support for text, image and video input. Sparse attention keeps long-context inference cheap enough that the API undercuts most closed frontier models by an order of magnitude. The M2.x line that preceded it remains the cheaper pick for coding agents, at roughly 10 billion active parameters.

• Best for: long documents, video and image analysis, and high-volume agent work where token cost dominates.
• Watch out: the licence is MiniMax's own rather than plain MIT or Apache, and the family updates every few months — pin a version.

How to choose between them

• Cheapest frontier reasoning: DeepSeek V4.1-Flash
• Self-hosting or fine-tuning on your own GPUs: Qwen3.8-27B
• A coding agent that has to output real Office files: GLM-5.3
• The hardest long-running engineering work, with vision: Kimi K3 (K2.6 if budget matters)
• Video and document input at volume: MiniMax M3

All five are reachable through an API, so that is where a trial belongs. Give two of them the same task from your own backlog, once, and judge the output the way you would judge a colleague's draft. The decision usually makes itself inside an hour.

What all five have in common

Three things apply across the list. Benchmark numbers come from the vendors and are run on the vendors' own harnesses, so treat a two-point gap as noise. Documentation is Chinese-first, and the English pages lag by weeks. And every API call leaves your network for a server in mainland China unless you self-host — which is the real reason Qwen's 27B is the most widely deployed model here, whatever the leaderboards say.