Top 5 Chinese Open-Source LLMs Dominating 2026

TL;DR: China now ships the strongest open-weight models in the world. Moonshot's Kimi K3 (2.8T parameters) is the largest open release ever made, DeepSeek V4-Pro matches closed frontier models on coding under an MIT licence, and Z.ai's GLM-5.2 leads open-weights coding benchmarks outright. The catch: licences vary wildly, and Alibaba has quietly moved its flagship tier behind an API.

Chinese open-source large language models stopped playing catch-up some time in the last twelve months. In July 2026, the best model you can download and run on your own hardware is Chinese — and depending on the benchmark, it is within a couple of points of the best model anyone can rent.

That shift is not subtle. China‘s labs released four frontier-class open-weight models in the first half of 2026 alone, and the newest of them — Kimi K3, whose weights landed on 27 July — is a 2.8-trillion-parameter model published as roughly 1.56 TB of downloadable shards.

What “open-source” means here

The July 2026 Ranking at a Glance

Rank	Model	Developer	Params	Context	Licence	Headline result
1	Kimi K3	Moonshot AI	2.8T (104B active)	1M	Kimi K3 (custom)	88.3 Terminal-Bench 2.1
2	DeepSeek V4-Pro	DeepSeek	1.6T (49B active)	1M	MIT	80.6 SWE-bench Verified
3	GLM-5.2	Z.ai (ex-Zhipu)	753B (~40B active)	1M	MIT	62.1 SWE-bench Pro
4	MiniMax M3	MiniMax	427B (~23B active)	1M	minimax-community	80.5 SWE-bench Verified
5	Qwen3.6-35B-A3B	Alibaba	35B (3B active)	256K	Apache 2.0	73.4 SWE-bench Verified

Note how tightly the top four cluster on SWE-bench Verified — 80.6, 80.5, 80.2, 77.8. That benchmark has effectively saturated for frontier open-weight models, which is why the ranking below leans on agentic benchmarks (Terminal-Bench, SWE-bench Pro) where the models still genuinely separate.

1. Kimi K3:

Developer: Moonshot AI (月之暗面)

Moonshot AI released Kimi K3 on 16 July 2026 and published the full weights eleven days later, on 27 July. At 2.8 trillion total parameters — 96 shards and roughly 1.56 TB on Hugging Face — it is the largest open-weight release in history by a wide margin. Our guide to every Kimi model traces the lineage from K1 to here.

The architecture is a sparse mixture-of-experts with 896 experts, of which 16 activate per token, giving 104 billion active parameters. Moonshot reports roughly 2.5× better overall scaling efficiency than Kimi K2.

Key Features

• 1,048,576-token context window — a genuine million tokens, not an extrapolated claim
• Native multimodality: text, images and video understood inside the same model
• Long-horizon coding: sustains multi-hour engineering sessions across large repositories with minimal oversight
• KDA & Gated MLA attention across 93 layers, with a 160K vocabulary

Why It Ranks #1

• 88.3 on Terminal-Bench 2.1 — ahead of Claude Fable 5 and Claude Opus 4.8, and within half a point of GPT-5.6 Sol
• 81.2 on FrontierSWE and 42.0 on SWE Marathon, a multi-hour whole-project benchmark where it leads every model tested
• #1 on the Frontend Code Arena at 1,679 Elo, and #3 overall on the Artificial Analysis Intelligence Index
• Official evaluation results on the model card: 93.5 GPQA Diamond, 67.5 DeepSWE, 56 on Humanity’s Last Exam

K3 is not the best at everything. It trails GPT-5.6 Sol on DeepSEE-style deep repository analysis (67.5 vs 73.0) and sits well behind Claude Fable 5 on Humanity’s Last Exam. It is a better agent than it is a pure reasoner — which, for teams building autonomous coding workflows, is the trade worth making.

2. DeepSeek V4:

Developer: DeepSeek (深度求索)

DeepSeek V4 shipped on 24 April 2026 in two sizes: V4-Pro at 1.6 trillion total parameters with 49 billion active, and V4-Flash at 284 billion total with 13 billion active. Both are MIT-licensed with no revenue thresholds, no attribution clauses and no regional restrictions — the most permissive terms of any model in this guide.

V4-Pro is the most downloaded model here by a factor of ten, with over 1.6 million downloads in a single month. It was pre-trained on more than 32 trillion tokens using the Muon optimiser, then post-trained through independent domain-expert cultivation followed by on-policy distillation into a single model.

Key Features

• 1M-token context with up to 384K tokens of output
• Hybrid attention combining Compressed Sparse Attention and Heavily Compressed Attention — 3.7× lower single-token FLOPs and 9.5× smaller KV cache than V3.2 at long context
• Trained on Huawei Ascend 950 chips and Cambricon accelerators rather than Nvidia GPUs
• The DeepSeek-Coder line, still MIT and still maintained, supports 338 programming languages

Benchmarks & Results

• 80.6 on SWE-bench Verified — statistically tied with Claude Opus 4.7 at 80.8
• 93.5 on LiveCodeBench, the highest figure in this guide
• Codeforces rating 3,206, ahead of GPT-5.5 at 3,168
• 67.9 on Terminal-Bench 2.0; 55.4 on SWE-bench Pro; 90.1 GPQA Diamond

The API is also the cheapest way to rent frontier-class output: $0.435 per million input tokens and $0.87 per million output, made permanent on 22 May 2026, with cache hits at $0.003625. If you want the best coding model you can legally do anything with, this is it. See also our breakdown of every DeepSeek model. For a deeper look at how it stacks up against Western models, see our comparison of Chinese AI coding assistants.

3. GLM-5.2:

Developer: Z.ai (formerly Zhipu AI, 智谱)

Zhipu AI now trades as Z.ai, and GLM-5.2 — released 13 June 2026 — is its flagship. It is a 753-billion-parameter mixture-of-experts model with roughly 40 billion active per token, a solid 1M-token context, and up to 131,072 tokens of output.

Z.ai describes the licensing position bluntly on the model card: “Pure Open: An MIT open-source licence — no regional limits, technical access without borders.” That framing is aimed squarely at the custom licences its competitors now use.

Key Features

• IndexShare architecture, reusing one indexer across every four sparse attention layers — 2.9× fewer per-token FLOPs at 1M context
• Two reasoning-effort levels (“thinking” and “max thinking”) to trade latency against accuracy
• Day-one support for eight third-party coding agents including Claude Code, Cline, Roo Code, OpenCode, Goose, Crush and Kilo Code
• Delivered through the GLM Coding Plan (Lite $18/mo, Pro $72, Max $160) or a metered API at $1.40 in / $4.40 out per million tokens

Benchmarks & Results

• 62.1 on SWE-bench Pro — ahead of GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2)
• 81.0 on Terminal-Bench 2.1, up from GLM-5.1’s 63.5 — the largest generation-over-generation jump in this guide
• 77.0 on MCP-Atlas for tool use; 63.7 on ProgramBench
• Predecessor GLM-5 scored 77.8 on SWE-bench Verified, trained on 28.5 trillion tokens across a 100,000-chip Huawei Ascend cluster

On 2 July 2026 Z.ai also launched ZCode, a free desktop agentic development environment built around GLM-5.2. If your requirement is “strongest coding model with zero licensing friction,” GLM-5.2 is the answer — and it also powers the tools in our guide to Chinese AI coding assistants.

4. MiniMax M3:

Developer: MiniMax (稀宇科技)

Released 1 June 2026, MiniMax M3 is a 427-billion-parameter mixture-of-experts model with roughly 23 billion active per token. It is the smallest model in the top four and, on SWE-bench Verified, gives up only a tenth of a point to DeepSeek’s 1.6-trillion-parameter flagship.

Key Features

• Native multimodality from step one — mixed-modality training across text, image and video rather than a bolted-on vision encoder
• MiniMax Sparse Attention (MSA): 9× faster prefill and 15× faster decode than M2 at 1M context, cutting per-token compute to one twentieth
• 1M-token context window with agent, coding and video capabilities in one checkpoint
• Best performance-per-parameter of any model here — the practical choice if you are GPU-constrained but still want frontier results

Benchmarks & Results

• 80.5 on SWE-bench Verified — matching Gemini 3.1 Pro (80.6) and beating Claude Sonnet 4.6 (79.6)
• 59.0 on SWE-bench Pro, ahead of GPT-5.5’s 58.6
• 66.0 on Terminal-Bench 2.1, essentially level with Claude Opus 4.7 at 66.1
• 83.5 on BrowseComp and 63.7 on SVG-Bench, reflecting the multimodal training

The catch is licensing. M3’s weights ship under a custom minimax-community licence, and commercial use of M3 or its derivatives requires a separate agreement with MiniMax. It is more permissive than M2.7’s terms, but it is not MIT — budget for a legal conversation if this is going into a product. MiniMax is profiled in our roundup of the hottest Chinese AI startups.

5. Qwen3.6-35B-A3B:

Developer: Alibaba Cloud (通义千问)

Qwen3.6-35B-A3B is a 35-billion-parameter mixture-of-experts model with roughly 3 billion active per token, released 16 April 2026 as the first open-weight variant of the Qwen3.6 series. It ranks fifth on raw capability and first on almost everything else that matters to a working team.

The number that tells the story is downloads: over 6.1 million in a single month, roughly four times DeepSeek V4-Pro’s and forty times MiniMax M3’s. With 222 adapters and hundreds of fine-tunes on Hugging Face, Qwen is where the open-source tooling ecosystem actually lives.

Key Features

• Apache 2.0 — the cleanest licence in this guide, with no revenue thresholds or attribution requirements
• Runs on a 24GB Mac with GGUF quantisation — the only model here you can realistically run on a laptop
• 73.4 on SWE-bench Verified, remarkable for 3B active parameters, and 20%+ ahead of Gemma4-31B on agentic coding
• Compatible out of the box with Transformers, vLLM, SGLang and KTransformers
• Over 200,000 derivative models across the wider Qwen family — the first open model family to pass that mark

How to Choose Your Ideal Chinese Open-Source LLM

Benchmark position is the least useful selection criterion once the top four sit within three points of each other. These are the questions that actually decide it.

1. What does your licence allow?

This is now the first question, not the fourth. DeepSeek V4 and GLM-5.2 are MIT — do anything. Qwen3.6 is Apache 2.0 — equally clean. Kimi K3 has revenue and attribution thresholds. MiniMax M3 requires a separate commercial agreement. If you are shipping a product, that ordering may decide the whole thing.

2. Can you actually serve it?

Kimi K3’s weights are about 1.56 TB. Serving a 2.8T-parameter MoE is a multi-node infrastructure project, not a docker run. MiniMax M3 at 427B is far more tractable, and Qwen3.6-35B-A3B runs on a single consumer GPU or a 24GB Mac. Match the model to the hardware you have, not the leaderboard you want to top.

3. Agentic work or single-shot generation?

For multi-step agent runs, Terminal-Bench 2.1 is the number to read: Kimi K3 (88.3) and GLM-5.2 (81.0) lead. For single-shot code repair and generation, SWE-bench Verified and LiveCodeBench matter more, and DeepSeek V4-Pro leads on both.

4. Do you need multimodality?

Kimi K3 and MiniMax M3 handle text, image and video natively. DeepSeek V4 and GLM-5.2 are text-first. If your pipeline involves screenshots, diagrams or video, that halves the shortlist immediately.

5. Self-host or API?

Calling a vendor API means your prompts transit servers governed by Chinese data law. For most projects that is a non-issue; for regulated workloads or customer code it is a blocker. The entire point of open weights is that you can avoid the question — every model here can run inside your own network.

6. How healthy is the ecosystem?

Quantisations, adapters, serving-stack support and community fine-tunes decide how fast you ship. On that axis Qwen wins outright despite ranking fifth on capability, with DeepSeek a clear second.

Honourable Mentions

Three more worth tracking, plus two names that dominated this list a year ago and no longer belong on it.

• Kimi K2.7-Code (12 June 2026) — Moonshot’s coding specialist on the K2.6 architecture, Modified MIT, with roughly 30% lower reasoning-token usage than K2.6 and 81.1% on MCPMark Verified
• Tencent Hunyuan — strong multimodal and 3D generation models, covered in our roundup of the top 10 Chinese AI companies
• MiniCPM (面壁智能) — still the reference point for genuinely small on-device models, where a 36B MoE is far too heavy

Two notable departures from the 2025 edition of this list: Yi 1.5 (01.AI) and Baichuan 4. Both were credible picks eighteen months ago, and neither has kept pace with the frontier open-weight releases of 2026. If you are running either in production, the models above will outperform them on essentially every axis — treat a migration as low-risk, high-return.

What Changed in the First Half of 2026

• Open weights reached the frontier. The gap between the best model you can download and the best model you can rent is now a few points on most coding benchmarks, and zero on some — we put hard numbers on that gap in US vs China AI models.
• 1M context became standard. Every model in the top four ships a million-token window. It is no longer a differentiator.
• Licences fragmented. As the models got good, the terms got specific. “Open weights” now spans MIT through to revenue-capped commercial agreements — and the distinction is a business decision, not a footnote.
• Domestic silicon stopped being a handicap. DeepSeek V4 trained on Huawei Ascend 950 and Cambricon hardware; GLM-5 on a 100,000-chip Ascend cluster. Export controls have not capped the frontier the way many expected.

The world of open-source AI is moving fast, and having the right people on your team makes the difference between reading about these models and running them.

Second Talent helps you hire AI developers, researchers and engineers who work with open-weight LLMs day to day — fine-tuning, quantising and serving them in production.

Get in touch to see how Second Talent can connect you with the right AI engineers for your stack.

FAQs About Chinese Open-Source LLMs

Q1: Can I use these models for commercial projects?
It depends on the model, and the answer changed in 2026. DeepSeek V4 and GLM-5.2 are MIT; Qwen3.6 open models are Apache 2.0 — all three allow unrestricted commercial use. Kimi K3 permits commercial sale but requires a separate agreement above $20M in model-as-a-service revenue and on-screen credit above 100M monthly users. MiniMax M3 requires a separate commercial licence outright. Always read the licence file in the repository.

Q2: Which model is best for coding tasks?
Kimi K3 for multi-step agentic work (88.3 Terminal-Bench 2.1). DeepSeek V4-Pro for single-shot code repair and generation (80.6 SWE-bench Verified, 93.5 LiveCodeBench). GLM-5.2 if you want the strongest coding model with a genuinely unrestricted licence.

Q3: What does “MoE” mean, and why is it important?
Mixture-of-Experts routes each token to a small subset of the network. Kimi K3 has 896 experts but activates only 16 per token, so a 2.8T-parameter model costs roughly 104B parameters’ worth of compute per token. It is what makes trillion-parameter models servable at all.

Q4: Are these models effective in English?
Yes. All five are trained bilingually and are competitive with Western models on English benchmarks — GPQA Diamond, LiveCodeBench and SWE-bench are English-language evaluations, and these models lead or match on them.

Q5: What if I have limited GPU resources?
Start with Qwen3.6-35B-A3B — 3B active parameters, GGUF quantisation, runs on a 24GB Mac. Next step up is DeepSeek V4-Flash (284B total, 13B active) or MiniMax M3 at 427B. Kimi K3 at 1.56 TB of weights is a multi-node deployment.

Q6: Is the best open-weight model now Chinese?
On coding and agentic benchmarks in July 2026, yes. Kimi K3, DeepSeek V4-Pro and GLM-5.2 lead the open-weight field, and the nearest Western open releases trail them on most published coding evaluations.