Kimi K3, Qwen3, GLM-4.6-who's actually #1 in China?

You can't rank them without a scenario. Long context and open-source size: K3. Coding and price-performance: Qwen3 Max / 235B. Agent and tool use: GLM-4.6. On Artificial Analysis's overall index, K3 sits at #3 (behind Fable 5 and GPT-5.6 Sol Max)-the highest among Chinese models-but on API price-performance alone, GLM and Qwen hit harder.

For long documents, Kimi K3 or Qwen3 Max?

Both reach 1M tokens, but K3 is natively 1M with thinking mode always on, so it handles ultra-long docs more decisively. Some Qwen3 Max versions cap at 256K; whether 1M is available depends on the exact variant. Budget ample, want the ceiling: K3. Want stability, cheapness, Alibaba Cloud ecosystem: Qwen3 Max.

Which open weights for self-hosting?

It depends on compute. K3 is a 2.8T MoE-forget a single GPU, you need an H100 cluster. Qwen3-235B-A22B fits on a single H100 and has the best price-performance. GLM-4.5 open variants are smaller, single-GPU-friendly, but a tier below flagship. If your compute isn't there, don't force K3.

Who's strongest at coding?

Arena's front-end coding blind test: Kimi K3 first. SWE-bench Verified public boards: Qwen3 Coder top among Chinese. GLM-4.6 is mid-upper and leans agent, not pure code. For front-end / full-stack code completion: K3 or Qwen3 Coder. For "agents that fix their own bugs": GLM.

For agents / tool use, who do you back?

GLM is the most turnkey in China-AutoGLM already ships browser / phone agents as products, native Chinese, no VPN, web-usable. K3 takes the "reasoning agent" route via thinking mode + long context. Qwen3 goes the Qwen-Agent framework route: flexible but you assemble it yourself. Product-grade, China-usable, fast: GLM.

Kimi K3 vs Qwen vs GLM: China's Big-Three LLM Showdown

China's LLM table has shifted over the past year from "who can catch up to GPT" to "who is number one in China." DeepSeek was the original breaker, but the three still holding seats in the second half of 2026 are Kimi, Qwen, and Zhipu GLM. This piece reviews them across six dimensions-long context, coding, agent/tool use, multimodal, pricing, and open source-and hands you a table you can pick from. Disclosure up front: benchmark numbers are from public leaderboards and third-party trackers, labeled as "representative figures, not self-tested"; prices are public API ranges, subject to each vendor's official page.

Why the "Big Three" Are Worth a Showdown

OpenAI is raising prices, Claude is throttling, Gemini is hit-or-miss, and data-export compliance keeps biting. So in 2026, Chinese developers have been quietly flipping "domestic alternative" from fallback to default. All three come from serious stock:

Kimi (Moonshot AI): launched the K3 API on July 16, then open-sourced the weights on the night of July 27. 2.8-trillion-parameter MoE, 1-million-token context-the largest open-source model in the world right now.
Qwen (Alibaba): the giant most committed to open source. The Qwen3 line releases weights across the board from 0.6B up to flagship, plus a closed Qwen3 Max as the API weapon.
Zhipu GLM (Tsinghua): the first Chinese LLM maker to commercialize at scale. GLM-4.5 is open, while closed GLM-4.6 + AutoGLM are the bets on agents.

Each walks a different road: Kimi bets on "long context + open weights for community," Qwen bets on "full-size open source + Alibaba Cloud backing," GLM bets on "agents + commercialize first." A showdown is how you see where each road actually leads.

Subjects and Method

Subjects (July 2026 versions):

Model	Flagship	One-line positioning
Kimi K3	K3 (open + API)	Long context + multimodal + open-source behemoth
Qwen Qwen3	Qwen3 Max (closed) / Qwen3-235B-A22B (open)	Full-size open source + Alibaba Cloud ecosystem
Zhipu GLM	GLM-4.6 (closed) / GLM-4.5 (open)	Agent bet + earliest to commercialize

Dimensions: long context, coding, agent/tool use, multimodal, pricing/API, open source. Each dimension gets a "best scenario" verdict. All benchmarks come from vendor-published figures or third-party leaderboards (Artificial Analysis, Arena, Vals AI, SWE-bench, LiveCodeBench, etc.) and are representative figures, not self-tested; pricing uses each vendor's official public ranges, marked "subject to the official page."

Core Capability Comparison

Dimension	Kimi K3	Qwen3 Max / 235B	Zhipu GLM-4.6 / 4.5
Total params	2.8T MoE (16/896 active)	235B MoE (22B active)	GLM-4.5 ~355B MoE (4.5 not fully disclosed)
Context window	1M tokens	1M tokens (some versions 256K)	200K (GLM-4.6) / 128K (GLM-4.5)
Native multimodal	Vision-native	Qwen3-VL (separate model)	GLM-4V (vision) / CogAgent
Thinking mode	Built-in, always-on	Optional / hybrid	Built-in (GLM-4.6)
Coding	Arena front-end coding blind test #1	Qwen3 Coder, purpose-built	Mid-upper, agent-leaning
Agent / tool use	Strong (thinking + long context)	Strong (Qwen-Agent framework)	Strongest (AutoGLM / CogAgent)
Open source	Full weights (July 27, MXFP4)	Full size range (HF / ModelScope)	GLM-4.5 series
Commercial license	Commercial OK, self-hostable	Commercial (Apache-2.0 / Qwen License)	Commercial (check terms)
Best for	Long docs / code / self-host flagship	Full-stack / Alibaba Cloud users / Chinese	Agent apps / domestic enterprises

Each Model, in Turn

Kimi K3: The Long-Context Open Bomb

K3 is Moonshot's counterpunch after 18 months under DeepSeek's thumb-2.8T params, 1M context, weights dumped on July 27. Two architectural highlights: Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals (a residual-connection replacement), both previously posted on GitHub. Weights ship in MXFP4 with MXFP8 activations, with quantization-aware training baked in, so the original runs don't feel "shoehorned in."

Strengths: long docs, end of story. 1M-token context plus an always-on thinking mode means you feed it an entire contract, an entire repo, dozens of pages of financials-without slicing, without RAG. It took first place on Arena's front-end coding blind test, beating top US models; Artificial Analysis ranks it third overall, behind Fable 5 and GPT-5.6 Sol Max.

Weaknesses: expensive and heavy. Self-hosting 2.8T params is wildly out of reach for a single GPU-you need an H100 cluster minimum. API pricing is flagship-tier, so heavy calls cost a notch more than mid-range Qwen. Multimodal only covers vision; no audio, no video.

Qwen3: The Giant Most Committed to Open Source

Alibaba actually puts money behind open source: Qwen3 drops weights across the whole size range from 0.6B to 235B, on both Hugging Face and ModelScope, with closed Qwen3 Max as the API flagship. The flagship MoE is Qwen3-235B-A22B (235B total, 22B active)-far denser than K3, so single-GPU inference is realistic.

Strengths: price-performance and full stack. A 0.6B runs on phones and embedded devices; the 235B flagship pushes on servers; one product line covers edge to cloud. Coding has the purpose-built Qwen3 Coder (consistently top among Chinese models on SWE-bench public boards), agents have the Qwen-Agent framework, multimodal has Qwen3-VL-each track actively maintained. Alibaba Cloud's Bailian (DashScope) API is the most stable in China and the most aggressive on price.

Weaknesses: flagship size still trails K3, so extreme long-context (>500K tokens) is less decisive than K3; the size range is wide enough that picking on first try is dizzying.

Zhipu GLM: The Heaviest Agent Bet

Zhipu is Tsinghua-spawned, the first Chinese LLM maker to commercialize and the first to push agents. GLM-4.5 is open, closed GLM-4.6 is the API flagship, AutoGLM turns browser and phone use into agents, and CogAgent specializes in GUI operations. Context sits at 200K-they didn't chase the million-token race.

Strengths: agents end to end. AutoGLM is a browser / phone agent, native Chinese, no VPN, usable from a web page-the lowest barrier for Chinese beginners. GLM-4.6's thinking mode and tool-call pipeline are the most polished; for "apps that call tools to finish tasks," GLM is the most turnkey of the three. Zhipu was also the earliest in China to mirror the OpenAI API shape, so migration cost is low.

Weaknesses: the shortest context, 200K, last of the three; vision goes through GLM-4V as a separate line, not native like K3; open-source GLM-4.5 sits a tier below K3 / Qwen3 flagship, so a self-hosted flagship means closed API only.

Pricing Comparison

Public API pricing ranges (RMB, per million tokens), subject to the official page, may change:

Model	Input price	Output price	Notes
Kimi K3 API	~10-20 yuan	~60-100 yuan	Flagship tier, Fable-5 class
Qwen3 Max	~5-15 yuan	~15-50 yuan	Alibaba Cloud Bailian, tiered
Qwen3-235B (open, self-host)	0 (compute only)	0 (compute only)	GPU cost only
Zhipu GLM-4.6	~5-10 yuan	~10-30 yuan	2025 price-war initiator
Zhipu GLM-4.5 Air (open, self-host)	0 (compute only)	0 (compute only)	GPU cost only

Takeaways:

On the self-host road, K3 has the biggest flagship size and highest compute cost; Qwen3-235B and GLM-4.5 fit on a single GPU, with friendlier price-performance.
On the API side, GLM-4.6 has the lowest historical pricing and started the 2025 price war; Qwen3 Max is the most stable mid-tier; K3 API is the most expensive but has the highest capability ceiling.
Pricing is a live war-what's written today may change tomorrow; always defer to the official page.

Representative Test Data

Aggregated from public leaderboards and third-party trackers, representative figures, not self-tested:

Test	Kimi K3	Qwen3 Max / 235B	GLM-4.6
Context window	1M tokens	1M tokens	200K tokens
Arena front-end coding blind test	#1	Top 3	Top 5
Artificial Analysis Intelligence Index	#3 (behind Fable 5, GPT-5.6 Sol Max)	Top 5	Top 8
SWE-bench Verified (public)	Mid-upper	Top among Chinese	Mid-upper
LiveCodeBench	Mid-upper	Top among Chinese	Mid-upper
Tool use / agent	Strong (thinking + long context)	Strong (Qwen-Agent)	Strongest (AutoGLM / CogAgent)
Multimodal	Vision-native	Qwen3-VL	GLM-4V
Open weights	2.8T MXFP4	Full size range	GLM-4.5 series

Takeaways:

For long docs, K3 and Qwen3 Max both reach 1M tokens; GLM sits at 200K-ultra-long context is the first two's home turf.
For coding, Qwen3 Coder sits consistently at the top of Chinese public boards; K3 took first on Arena's front-end blind test; GLM is mid-upper but agent-leaning.
For agents / tool use, GLM bets hardest: AutoGLM already ships browser / phone agents as products; K3 follows via thinking mode + long context; Qwen3 takes the framework route (Qwen-Agent).
For multimodal, all three only cover vision-K3 natively (no model switching), while Qwen3-VL and GLM-4V are separate models. Speech and video are not in these three's wheelhouse yet.

Why No DeepSeek

A common question: DeepSeek is also in China's top tier, so why isn't it here? Three reasons.

First, DeepSeek's trajectory differs from these three. DeepSeek V4 Pro (~1.6T params) was the first to push Chinese open source into global view in 2024, and Artificial Analysis consistently ranks it top five overall. But its rhythm is more "research-style open source": publish papers, drop weights, slowly iterate the commercial API. It hasn't pushed a "commercial flagship + open-source flagship" double line to the front of the market the way Kimi/K3, Qwen3, and GLM-4.6 have.

Second, all three are running a "commercial API + open weights" double line; DeepSeek's commercial API is slower, more "weights-as-product," so the comparable dimensions differ.

Third, the second tier: MiniMax (betting on audio/video), Baichuan, Stepfun, 01.AI, etc. Each has a specialty, but on overall capability, API ecosystem, and developer base, none has reached these three's level. So "Big Three" is essentially the current market consensus.

If you insist on including DeepSeek, that makes it a "Big Four"-DeepSeek V4 Pro is top-tier on long context (also ~1M class), coding, and open-source size, with aggressive pricing. But on "commercial-API readiness + agent ecosystem + application layer," it is genuinely on a different road.

Selection Advice

Processing entire contracts / repos / dozens of financial pages -> Kimi K3; 1M context + thinking mode settle it, with self-hosting optional.
Price-performance + full-size open source + Alibaba Cloud ecosystem -> Qwen3; one line from 0.6B to 235B, edge to cloud.
Building "agent apps that call tools themselves," especially browser / phone operations -> Zhipu GLM-4.6 + AutoGLM, the most turnkey domestic option.
Budget-sensitive small-to-mid teams -> Qwen3 Max API or GLM-4.6, the cheapest; for heavy calls, consider self-hosted Qwen3-235B.
Budget ample, want the capability ceiling -> Kimi K3 API.
Private deployment + flagship size -> Only Kimi K3 (GLM-4.6 / Qwen3 Max are closed); compute bar is an H100 cluster.
Private deployment + price-performance -> Qwen3-235B-A22B (open MoE) or GLM-4.5 (open), single-GPU-runnable.

Three Pitfalls

Don't pit the "Big Three" head-on against Fable 5 / GPT-5.6 flagship reasoning: complex long-chain reasoning, brutal math-domestic flagships still trail the international top tier by a notch. The gap has compressed from over a year a year ago to three to five months now (Nathan Lambert's read), but hasn't closed. Don't oversell.
Long context is not a free lunch: 1M tokens sounds great, but real inference cost, first-token latency, and the "lost in the middle" problem are real prices. Decide whether your scenario truly needs >200K tokens before reaching for a flagship.
Open source isn't free for commercial use: K3 / Qwen3 / GLM-4.5 all grant commercial licenses, but the terms differ (Apache-2.0, Qwen License, vendor-own licenses). Read each clause before commercial use-especially around distillation, redistribution, and compliance notices.

References

Kimi K3 open weights download and API (Moonshot AI): platform.moonshot.cn
Qwen3 open-source repository: github.com/QwenLM/Qwen3
Alibaba Cloud Bailian (DashScope) docs: help.aliyun.com/zh/model-studio/
Zhipu GLM open-source repository (Zhipu AI): github.com/zhipuai
Zhipu open platform bigmodel.cn: bigmodel.cn
Artificial Analysis model leaderboard: artificialanalysis.ai