Which China AI foundation models should builders evaluate first?

Start with the current watchlist rather than a single benchmark winner. Qwen remains the baseline open-weight family to evaluate first. DeepSeek remains the baseline hosted API and open-weight family for cost-sensitive reasoning and coding workloads. GLM is back on the short list because Z.AI now emphasizes GLM-4.5 and GLM-5.1 for agent and coding workflows. Kimi matters for long-context and agent products. MiniMax matters when multimodal and developer-platform packaging are part of the decision. ERNIE matters mainly when Baidu ecosystem or enterprise Chinese-language deployment is the real constraint.

How does Qwen compare to DeepSeek for commercial use?

Use current official surfaces, not frozen comparison tables. Qwen remains the cleaner open-weight route for multilingual general-purpose work because the official Qwen ecosystem still centers Apache 2.0 releases and broad tooling support. DeepSeek remains the stronger route to watch when hosted API price-performance, coding, and reasoning matter. The right practical choice depends on whether your bottleneck is open-weight flexibility, hosted API economics, or coding-agent behavior.

Which Chinese foundation models are available outside China?

The safe answer is to verify each provider at its current official surface. Qwen and DeepSeek clearly expose public global-facing release surfaces. Kimi, GLM, and MiniMax now also publish builder-facing hosted access pages. ERNIE remains the one to treat as more enterprise-gated and verification-heavy. When in doubt, trust the current official access page over any older summary.

What is the best open-source Chinese LLM for fine-tuning?

Qwen remains the safest default family to fine-tune because the open-weight release path, license clarity, and ecosystem tooling are still strongest there. DeepSeek-distilled variants remain important for reasoning-heavy fine-tuning. The right answer should still be verified at the current official model card and tooling support surface before you standardize internally.

How do Chinese foundation models compare to GPT-4o on benchmarks?

The more durable answer is that China AI foundation models are now strong enough that builders should compare by workload, not by one static benchmark page. Qwen and DeepSeek belong in almost every serious evaluation set. GLM and Kimi belong there when agent, coding, or long-context behavior matter. MiniMax belongs there when multimodal product packaging matters. Use the official release page and current model card to validate any specific benchmark claim before acting on it.

Where can I access Kimi, Qwen, and DeepSeek APIs in English?

Use the current official docs for each provider. DeepSeek, Qwen, Kimi, GLM, and MiniMax all currently expose developer-facing hosted surfaces. The exact active model names, compatibility mode, and billing path should be verified there rather than copied from a stale comparison chart.

China AI Foundation Models: Qwen, DeepSeek, Kimi, GLM (2026)

Thesis

The most useful 2026 foundation-model view is a current watchlist, not a frozen benchmark table. Qwen remains the baseline open-weight family for multilingual and general-purpose builder work. DeepSeek remains the baseline hosted API and open-weight family for cost-sensitive reasoning and coding workloads. Z.AI's current docs now make GLM-4.5 and GLM-5.1 relevant again for agent and coding evaluation. Kimi matters for long-context and agent workflows. MiniMax matters when multimodal or product-surface considerations are part of the decision. ERNIE still matters mainly in Baidu-ecosystem or enterprise Chinese-language contexts.

Decision in 20 seconds

Use case	Recommended model	Why
Commercial fine-tuning (open license)	Qwen3-7B / Qwen3-14B (Apache 2.0)	Broadest tooling support; explicit commercial fine-tuning rights; 0.6B–235B family consistency
RAG over multilingual documents	Qwen3-32B or Qwen3-30B-A3B MoE	Best multilingual retrieval-augmented performance in the open-weight class
Long-document processing (100K+ tokens)	Kimi k2.5 / k2.6 (API via platform.moonshot.cn)	128K context native; competitive on long-context RULER benchmark
Code generation and STEM reasoning	DeepSeek-V3 (MIT)	HumanEval 84.2; MATH 81.6; MIT license; Together AI / Fireworks hosting available
Enterprise Chinese NLP (on-prem required)	ERNIE 4.0 (Baidu Cloud) or GLM-5 / GLM-5.1 (open.bigmodel.cn)	Best Chinese-language domain adaptation; ERNIE has strongest Baidu ecosystem integration

China AI foundation models — full comparison (2026)

Family	Current official status	Core facts to know now	License / access	Best fit
Qwen family	Official Qwen surfaces now span the Qwen3 baseline plus newer Qwen 3.6 coding-oriented branches.	Qwen remains the most important open-weight multilingual family, and its public release surfaces are still the cleanest English-first path for builders.	Apache 2.0 on the main open-weight line; hosted access through official Qwen and Model Studio surfaces.	Default family for open-weight evaluation, multilingual use, and hosted/open crossover.
DeepSeek family	The official docs now center the DeepSeek V4 hosted line rather than the older V3-only framing.	DeepSeek now needs to be read as both a hosted API story and an open-weight story; current pricing, context, and active model names should start from the official docs.	Use the current official docs and pricing pages for hosted access; use official release/model-card surfaces for the open-weight line.	Strongest route when coding, reasoning, and hosted API economics all matter.
Kimi family	The current Kimi platform foregrounds `kimi-k2.6`, `kimi-k2.5`, and Moonshot V1.	Kimi is now a long-context and agent product surface, not just a single benchmark story.	Hosted API route via the Moonshot/Kimi platform; verify current pricing and active branch on the live platform page.	Best when long-context and agent product behavior are the real selection criteria.
GLM family	Z.AI now positions GLM-5.1 as the flagship and GLM-5 as the main active family around it.	GLM should be back in the core watchlist because the current official docs explicitly push long-horizon coding, agent execution, and 200K context.	Check each branch separately on the official docs, because some routes are hosted-first and others are tied to older open-weight paths.	Best when coding agents, long-horizon execution, and commercial API comparison matter.
ERNIE family	Still primarily an enterprise and Baidu-ecosystem route rather than the easiest public global builder path.	ERNIE matters when enterprise Chinese-language deployment and Baidu cloud integration matter more than open-weight flexibility.	Use current Qianfan / Baidu Cloud docs as the source of truth for access and model path.	Best when enterprise Chinese NLP and Baidu ecosystem fit are the main constraint.
MiniMax family	The public MiniMax platform foregrounds developer plans, multimodal positioning, and public API pricing.	MiniMax now matters more as a multimodal product/platform surface than as one frozen text-model comparison row.	Verify active branch, pricing, and modality support on the official platform and pricing pages.	Best when multimodal product packaging or platform ergonomics are real selection factors.

API access methods — international builder guide

Access method	Models available	Notes	Best for
Direct API (lab)	Qwen (dashscope.aliyun.com), DeepSeek (platform.deepseek.com), Kimi (platform.moonshot.cn), MiniMax (minimax.io)	OpenAI-compatible /v1/chat/completions on all four; Stripe/international card accepted on DeepSeek, Kimi, MiniMax	Lowest latency; most up-to-date model versions
HuggingFace (weights)	Qwen3 (all sizes), DeepSeek-V3/R1, GLM-4-9B or newer GLM open weights, MiniMax-Text-01	Free download; Apache 2.0 / MIT; no rate limits once downloaded; requires GPU infrastructure	Self-hosted fine-tuning, private deployment, offline inference
International cloud hosting	Qwen3 (AWS Bedrock, Cloudflare Workers AI), DeepSeek (Together AI, Fireworks AI, Amazon Bedrock), GLM-4 (Replicate)	US-billing, no China account required; Together AI and Fireworks add <50ms latency overhead; pricing competitive with direct API	Teams without China payment methods or needing US data residency
Self-deployment (vLLM / Ollama)	Qwen3-0.6B to Qwen3-32B; DeepSeek-R1-Distill-Qwen-7B; GLM-4-9B or newer GLM open weights	vLLM 0.4+ supports Qwen3 natively; Ollama library has quantized versions for M-series Mac; A100 40GB handles Qwen3-30B-A3B MoE at full precision	Air-gapped environments, on-prem enterprise, zero API cost at scale

FAQ

What are the best China AI foundation models in 2026?: Use a current watchlist rather than a frozen best-model answer: Qwen and DeepSeek belong in almost every serious evaluation set, GLM and Kimi matter when agent/coding or long-context behavior matters, MiniMax matters when multimodal product packaging matters, and ERNIE matters when Baidu ecosystem or enterprise Chinese-language deployment is the real constraint.
How does Qwen compare to DeepSeek for commercial use?: Use the current official release surface before making a hard comparison. Qwen remains the cleaner open-weight route for multilingual and general-purpose builder work, while DeepSeek remains the route to watch when hosted API economics, coding, and reasoning matter most.
Which Chinese foundation models are available outside China?: Qwen3 and DeepSeek-V3/R1 via HuggingFace globally. Kimi API (platform.moonshot.cn) and MiniMax API (minimax.io) accept international payment. ERNIE 4.0 is primarily China-gated via Baidu Cloud. GLM-4-9B or newer GLM open weights are available; full GLM-5.1 API is at open.bigmodel.cn with international tier.
What is the best open-source Chinese LLM for fine-tuning?: Qwen3-7B and Qwen3-14B — Apache 2.0 is explicit on commercial fine-tuning rights, tooling (LLaMA-Factory, Axolotl, Unsloth, vLLM) is comprehensive, and the 0.6B–235B family shares a consistent architecture and tokenizer. DeepSeek-R1-Distill-Qwen-7B is an alternative for reasoning-heavy tasks.
How do Chinese foundation models compare to GPT-4o on benchmarks?: The practical answer is no longer one static row-by-row benchmark claim. These families are now strong enough that the real decision should happen by workload, access path, deployment style, and license fit. Use the current official release page or model card for any exact benchmark citation.
Where can I access Kimi, Qwen, and DeepSeek APIs in English?: Kimi: platform.moonshot.cn (English docs, international card). Qwen: dashscope.aliyun.com or AWS Bedrock (no China account needed). DeepSeek: platform.deepseek.com or Together AI / Fireworks AI (US billing, OpenAI-compatible endpoints on all three).

Companion pages in this cluster

If your question is about…	Go to	What's there
Which Chinese models are open-source and commercially usable	China AI Open Source Models	Full license comparison table — Apache 2.0, MIT, custom; download links and commercial use restrictions
Recent model releases and capability milestones	Model Release Tracker	Qwen3, DeepSeek, Kimi release timeline with benchmark data and source verification
Weekly digest of what changed in China AI	China AI Updates	Weekly signal digest — model releases, funding, policy, curated for builders
Which China AI companies to watch	Best China AI Companies	15-company shortlist: foundation model labs, workflow infra, physical AI — with monitoring frequencies
China AI startup funding rounds and IPOs	China AI Startup Funding Tracker	Funding tracking logic by company category and current market surface

Quotable summary: In 2026, the correct question is no longer "can Chinese foundation models match GPT-4o?" — they can, on most benchmarks. The question is which open license fits your deployment, which API fits your payment infrastructure, and whether you need on-prem weights or managed inference. Qwen3 and DeepSeek-V3 answer both license and capability simultaneously; Kimi fills the long-context gap; ERNIE and GLM-5/5.1 serve enterprise Chinese NLP. Pick by use case, not by country of origin.