Verdict: Yes, serious AI users should test Chinese models — but you need to make separate decisions about the task, the model, and the deployment path. DeepSeek V4 Pro is the best value for high-volume, price-sensitive API work. GLM-5.2 is the strongest open-weight option for long-horizon coding under a true MIT license. Kimi K3 commands premium pricing for near-frontier agentic performance. Qwen and MiniMax M3 fill niches in local deployment and multimodal workflows respectively. None of them is a universal replacement for a US frontier model, and none of them is interchangeable with another Chinese model.
Last verified: 2026-07-29
- DeepSeek V4 Pro: $0.435/$0.87 per M tokens, 1.6T params, 1M context, most capable Chinese model per NIST CAISI
- Kimi K3: $3/$15 per M tokens, 2.8T params, 1M context, launched July 16 2026, weights promised by July 27
- GLM-5.2: $1.40/$4.40 per M tokens, ~753B params, MIT license, #1 open-weight model on Artificial Analysis
- MiniMax M3: $0.30/$1.20 per M tokens, 428B params, custom community license, native multimodal
- Qwen 3.8: family spanning small open models to closed hosted Max line
- Volatile facts: Pricing, limits, and model versions change often. Re-check before committing.
Why "Chinese AI models" is a misleading label
The biggest mistake in the Chinese model conversation is treating an entire country's output as one product category. We use "Chinese models" as shorthand for cheap, or open, or local — and sometimes all three are true, and sometimes none of them are.
Consider what these five labs are actually doing:
- DeepSeek is pursuing extreme inference economics. V4 Pro costs 87 cents per million output tokens — roughly 34x less than GPT-5.6 Sol's $30 per million. It is a price-floor play.
- Moonshot AI is pricing Kimi K3 as a premium frontier product at $15 per million output tokens — the same tier as Anthropic's Sonnet. This is not a budget model. It is positioned as a flagship.
- Z.ai (Zhipu AI) ships GLM-5.2 under a true MIT license with no commercial strings, targeting long-horizon coding workloads where it reportedly beats GPT-5.5 on selected benchmarks.
- Alibaba's Qwen is a broad family that runs from genuinely small open models you can run on a laptop to a separate hosted Max line whose weights are closed. Qwen 3.8 spans the full spectrum.
- MiniMax combines coding, computer use, and image/video understanding in M3 under a custom community license that requires attribution and prohibits military use.
They do not share a price. They do not share a deployment path or a hardware burden. They do not even share a best use case. They just share a country of origin. Start there.
How far behind the US frontier are Chinese models, really?
In May 2026, the US government's Center for AI Standards and Innovation (CAISI) — a unit within NIST — published an evaluation of DeepSeek V4 Pro that provides one of the first formal, third-party measurements of the US-China AI capability gap.
Key findings:
- DeepSeek V4 Pro is the most capable PRC AI model evaluated by CAISI to date, across five domains: cyber, software engineering, natural sciences, abstract reasoning, and mathematics.
- Its capabilities lag behind the US frontier by approximately 8 months — meaning it performs similarly to GPT-5, which was released about 8 months earlier, even though DeepSeek's own self-reported evaluations claim parity with the more recent GPT-5.4 and Opus 4.6.
- The gap is smaller on individual tasks: the model is "spiky," with reinforcement-learned strengths in specific areas like mathematics (OTIS-AIME-2025, PUMaC 2024) and software engineering.
- DeepSeek V4 Pro is more cost-efficient than comparable US models on 5 of 7 benchmarks, ranging from 53% less expensive to 41% more expensive per correctly solved task compared to GPT-5.4 mini.
That last point is the one most people miss. Token price and finished-work cost can point in opposite directions. A model that costs 53% less per token may end up 41% more expensive for your specific workload if it needs more tokens to reach a correct answer, makes unnecessary tool calls, fails late, or hands you more cleanup. Cost per accepted result — not cost per token — is the gold standard. You have to define your task clearly enough to measure it.
The five Chinese frontier models compared
| Model | Total params | Active params | Context | Input $/M | Output $/M | License | Weights | Best for |
|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 Pro | 1.6T | 49B | 1M | $0.435 | $0.87 | MIT | Yes | High-volume API work, extraction, classification |
| Kimi K3 | 2.8T | ~16B (of 896 experts) | 1M | $3.00 | $15.00 | TBD (promised) | Not yet (promised Jul 27) | Long-horizon agentic coding, vision tasks |
| GLM-5.2 | ~753B | ~40B | 1M | $1.40 | $4.40 | MIT | Yes | Long-horizon coding, self-hosting with clean license |
| MiniMax M3 | 428B | ~23B | 1M | $0.30 | $1.20 | Custom (community) | Yes | Multimodal workflows, screenshot-to-code, UI agents |
| Qwen 3.8 | Varies | Varies | Varies | Varies | Varies | Apache 2.0 (open tiers) | Yes (open tiers) | Local deployment, small-model use cases, agent swarms |
Sources: NIST CAISI evaluation (May 2026); BenchLM pricing (July 2026); OpenRouter; Moonshot platform docs (July 16, 2026); Artificial Analysis (July 2026); Z.ai Hugging Face.
Pricing as of July 29, 2026. DeepSeek's 75% launch discount was made permanent on May 22, 2026. MiniMax lists "Permanent 50% off" rates. All pricing is subject to change.
Which Chinese model should you use for which job?
High-volume, price-sensitive API work: DeepSeek
If you are tackling extraction, classification, first-pass research, test generation, or any other reviewable work at volume, DeepSeek V4 Pro belongs in your test set. At $0.435/$0.87 per million tokens, it can change the economics of document processing, code generation, and research pipelines by 15x to 34x compared to US frontier APIs.
The key constraint: this is work where a machine can afford to check and recheck. If one ambiguous answer drives an unrecoverable action — a decision you cannot easily verify — you want to think hard about whether you are trading down quality in a way you cannot accept.
Long-horizon coding and research: GLM-5.2 and Kimi K3
For long-horizon coding, research, and agent work, test GLM-5.2, Kimi K3, Qwen, and MiniMax against the best US frontier models using the same tools and the same acceptance standard. Do not change your bar.
GLM-5.2 on coding is particularly strong — it tops the Artificial Analysis open-weights Intelligence Index at 51, and VentureBeat reports it beats GPT-5.5 on several long-horizon coding benchmarks at roughly one-sixth the cost. But that does not mean it is automatically a universal replacement. A model can write an excellent patch and still lose a repository-level task because it used the wrong tool, missed a test, or failed after 20 minutes of agent runtime.
Kimi K3 is Moonshot's bet on the premium tier: 2.8 trillion parameters, always-on reasoning, native vision, and $15 per million output tokens. On vendor-reported agentic benchmarks, it scores 88.3 on Terminal-Bench 2.0 and 81.2 on FrontierSWE. But those are self-reported numbers, and the always-on thinking mode means output-token bills are structurally high.
Local deployment: smaller Qwen models and DeepSeek R1 distills
If you are on personal hardware, look at the smaller Qwen models and the models DeepSeek distilled from R1. These are the releases behind the real local AI story. They are not the same artifact as full DeepSeek R1 (671B total parameters) or DeepSeek V4 Pro (1.6 trillion). DeepSeek's own repository documents that R1-distilled Qwen student models were fine-tuned on 800,000 samples curated from R1, with checkpoints at 1.5B, 7B, 14B, and 32B under MIT and Apache 2.0 licenses.
Local models make sense for private notes, offline work, predictable availability, and workflows where you would rather trade some capability for a high degree of control and privacy. The question is not whether a small Qwen model compares to Kimi K3. It is whether the small model is good enough for the exact job you want to keep on your machine.
For a deeper framework on building a local AI stack, see our 5-layer local AI stack guide and our guide on how to run Hermes Agent for free with Hunyuan 3.
Multimodal and UI agent workflows: MiniMax M3
MiniMax M3 is the first open-weight model to pair frontier agentic coding (80.5% SWE-bench Verified) with native multimodality — text, image, and video input from step zero of training. For UI automation and screenshot-to-code at low cost, M3 stands alone in the open-weight tier. The catch: its custom community license requires attribution and prior written authorization above $20 million in annual revenue, and prohibits military use. Read the license before building on it.
What does "open" actually mean for each model?
"Open" is not one property. You need to ask a series of separate questions:
- Can I call the model through an API? All five offer first-party APIs, but the provider controls the service, price, logging, and availability.
- Can I download the weights? GLM-5.2 and MiniMax M3: yes, on Hugging Face right now. DeepSeek V4 Pro: yes. Qwen open tiers: yes. Kimi K3: not yet, promised by July 27, 2026.
- Does the license permit my use? GLM-5.2 is true MIT — no strings. DeepSeek is MIT. Qwen open tiers are Apache 2.0. MiniMax M3 is a custom community license with revenue thresholds and military-use prohibitions. Kimi K3's license is TBD.
- Can my hardware run it at the context length I need at the speed I need? A model can be open on one dimension and closed — or simply impractical — on the rest. GLM-5.2's BF16 checkpoint is roughly 1.5 terabytes. You are firmly in multi-GPU territory. Downloading the weights does not magically solve the hardware and team requirements.
Here is the thing that gives US frontier models their advantage: you do not have to think about any of this if you are just signing up for OpenAI or Anthropic. That simplicity is a feature, not a bug.
Why enormous models can still be cheap: Mixture of Experts
One technical reason these enormous models can still be economical is Mixture of Experts (MoE) architecture. DeepSeek V4 Pro has 1.6 trillion total parameters but only 49 billion active per token. MiniMax M3 has 428 billion total but just 23 billion active. The router does not use the whole network for every token — part of the magic is in picking the right parameters to activate for a given token.
Active parameters help explain compute per token. Total parameters still determine storage, memory, networking, and serving burden. That is why a model can be inexpensive through an optimized API and absolutely absurd to run on your laptop. Do not turn the active count into a simple hardware estimate. You have to ask about checkpoint size, precision, context requirements, and serving topology.
If all of that sounds complicated, it is. And that is exactly the advantage frontier models have in the US — you do not have to think about it.
The distillation controversy: what we know and what we do not
Distillation is the technique at the center of the US-China AI competition, and it connects every major story of 2026. In the normal version, a powerful teacher model generates examples, reasoning traces, or preferences. A student model trains on those outputs and inherits some of the teacher's behavior without reproducing the full training program. This is standard practice. OpenAI sells distillation as a product. Frontier labs use it internally. DeepSeek openly used R1-generated samples to train smaller models based on Qwen and Llama — 800,000 samples, documented in the repository, MIT licensed. That is disclosed, legitimate teacher-to-student training.
The controversy begins when the teacher's outputs were never authorized for that use case.
What Anthropic alleges
In February 2026, Anthropic published a report accusing DeepSeek, Moonshot AI, and MiniMax of running what it called industrial-scale distillation campaigns against Claude. The numbers:
- ~24,000 fraudulent accounts generating more than 16 million exchanges with Claude.
- MiniMax alone drove over 13 million exchanges — more than three-quarters of the total — focused on agentic coding and tool orchestration. Anthropic said it detected the campaign while it was still active, "before MiniMax released the model it was training."
- Moonshot generated over 3.4 million exchanges targeting agentic reasoning, tool use, coding, and computer-use agent development. Anthropic said it attributed the campaign "with high confidence" through API request metadata that matched the public profiles of senior Moonshot staff.
- Anthropic does not sell Claude commercially in China, so by its account every one of the 24,000 accounts violated terms of service.
These are allegations, not court findings. The accused labs have not formally responded on the record.
The White House escalates
In April 2026, the White House OSTP released a policy memorandum accusing China of "industrial-scale" distillation of US AI models. Director Michael Kratsios said the US "has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI." The memo committed the government to sharing intelligence with US AI companies and exploring accountability measures. It did not name specific labs or validate Anthropic's numbers at the time.
That changed on July 22, 2026, when Kratsios publicly accused Moonshot AI by name, stating: "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." He alleged Moonshot built a "sophisticated internal platform" for large-scale covert distillation and accessed banned Nvidia GB300 chips through Thailand. Treasury Secretary Bessent warned sanctions may follow.
Why this matters even if you reserve judgment
The distinction that matters is authorization and conduct. A lab can distill its own model or a licensed teacher. It can train on API outputs contrary to contractual restrictions. It can use fake accounts and access-control evasion at scale. The technical method may overlap, but the contractual and potential legal claims differ.
Chips are physical. Model outputs are information. Restricting a teacher model's GPU does not control everything that model can teach through an API. Capability can move across the global internet through synthetic data, fine-tuning, reinforcement learning, and distillation. The transfer may not be perfect, but it may not need to be — it can reproduce enough valuable behavior fast enough and cheaply enough to be worth doing.
This means today's capability gap can close faster than hardware policy alone would suggest, and it gives frontier providers strong incentives to tighten access, monitoring, and account controls. For customers, that increases the value of owning your own evals and preserving an exit path. A workflow is much less exposed to one provider when the prompts, tests, and tools can be lifted and shifted to another model.
How to decide: a four-step evaluation framework
Step 1: Name the job
Is it bounded volume work? Local assistance? Sensitive internal work? Frontier judgment? Do not choose a model before you choose what you can tolerate from a failure perspective. The job determines the risk profile.
Step 2: Name the artifact and deployment
Is it an API? Downloadable weights? A promised future release? What license applies? What does the full model require at a realistic context length for your work? If you are looking at a first-party API, the provider controls the service, price, logging, and availability. A third-party host can run open weights in a region and contract any way you prefer. Self-hosting gives you the most control but makes you responsible for hardware, security, updates, and monitoring.
For a structured approach to testing multiple models, see our guide on how to test frontier AI models side by side.
Step 3: Measure cost per accepted result
Include input tokens, output tokens, reasoning traces, tool calls, retries, and latency — the entire chain, plus infrastructure if you are self-hosting. Do not let hidden costs accumulate. A cheap model can become expensive when it produces longer reasoning traces (which is what happens with always-on thinking models like Kimi K3). A premium model can end up being cheaper if it succeeds in one pass. For more on reasoning efficiency and frontier model competition, see our Claude Opus 5 review and the AI infrastructure supercycle analysis.
Step 4: Trace the data and the exit path
Identify which company receives the prompt, where it is stored and processed, the governing law and contract, and the retention and training terms. Then ask: can I move the workflow and keep my prompts, tools, and evaluations if the provider changes its price or policy?
DeepSeek's first-party privacy policy says personal data is processed and stored in China and inputs may be used to improve its technology. Alibaba says Model Studio customer data is not used for model training and offers deployment scopes that can exclude mainland China. An American third-party host running Qwen weights creates yet another path under American jurisdiction. A properly configured self-hosted model can keep prompts away from the original lab and remove provider-side runtime filters — but it does not remove refusal behavior or political bias learned in the weights. You may need to fine-tune.
When should you self-host vs. use an API?
Self-hosting only makes sense when four things are true:
- The weights are available and the license permits your use. GLM-5.2 and DeepSeek: yes, MIT. MiniMax M3: yes, but read the community license. Kimi K3: not yet.
- The data or sovereignty requirement is real enough that you need the model on your premises. "Real" means a compliance mandate, a contracts clause, or a genuine security posture — not a vague preference.
- The workload justifies the infrastructure. Running a 1.5-terabyte checkpoint across multiple GPUs with an MoE-aware scheduler, expert parallelism, and enough headroom for concurrent requests is a systems project, not a download.
- You have a team that can operate the stack safely. Self-hosting needs a named owner for authentication, model server patches, capacity planning, and observability. Without that accountability, you are converting vendor risk into an operating problem you may only discover when something goes wrong.
If those conditions are not present, look at an API or a managed third-party deployment. That will usually be a better solution.
For a framework on when to use managed services versus self-hosted infrastructure, see our guide on agent OS vs. agent framework and the AI agent loops engineering guide.
Should you use Chinese models in an agent swarm?
Yes — selectively. In testing for multi-agent workflows, Qwen proved useful as a swarm participant: a cheaper model doing grunt work while a frontier model orchestrates. The pattern is to let the expensive model decide and the cheap model execute, which delivers frontier-level intelligence at a fraction of the cost.
The key is that you can configure your agent framework to use whichever models you want. If you are building an agent pipeline, the same principles apply: test which Chinese model fits which role in the swarm, measure cost per accepted result for the whole chain, and make sure you can swap models without rewriting your workflow. For more on when multi-agent pipelines make sense — and when they do not — see our guide on single-agent vs. multi-agent architecture.
Kimi K3 also deserves a look here. Moonshot's launch positioned it as the #1 model on the Arena WebDev leaderboard ahead of both Fable 5 and GPT-5.6 Sol, with a 76% win rate on the Frontend Code Arena. If those numbers hold under independent testing, K3 could be a strong orchestrator model. But always treat vendor-reported benchmarks as signals to test, not facts to build on. For more on Kimi K3 specifically, see our coverage of the Kimi K3 open-weights drop and its 51% hallucination rate.
What this means for you
Chinese models are not a monolith. They are five (or more) entirely different products with different strategies, licenses, prices, and deployment paths. Country of origin is the beginning of due diligence, not the end.
- If you run high-volume, reviewable work, test DeepSeek V4 Pro. The economics can be transformative.
- If you want long-horizon coding with a clean MIT license, test GLM-5.2. It is the strongest open-weight option in that niche.
- If you need multimodal workflows (image, video, screenshot-to-code), test MiniMax M3. Read the community license first.
- If you build agent swarms, test Qwen as a participant and Kimi K3 as an orchestrator — but verify with your own evals.
- If the work involves sensitive data or unrecoverable decisions, trace the data path before you choose the model. A Chinese model on your own server is a different risk profile from the same model through a first-party API in China.
The honest caveat: benchmarks are easy to game and hard to trust. The CAISI evaluation found that DeepSeek V4 Pro scores better on its own self-reported evaluations than on independent CAISI evaluations. Always test on your own workloads with your own acceptance criteria. If you are not ready to go to that level of detail, you are probably not ready for a Chinese model conversation.
FAQ
Q: Are Chinese AI models caught up with US frontier models?
A: Not quite. NIST's CAISI evaluation (May 2026) found DeepSeek V4 Pro — the most capable Chinese model it tested — trails the US frontier by approximately 8 months. The gap is smaller on individual tasks where the model is "spiky" (strong in math and software engineering), but no Chinese model has matched the overall aggregate capability of GPT-5.5 or Claude Opus 4.8 as of July 2026. (Source: NIST CAISI)
Q: Is it legal to use Chinese AI models?
A: Using a Chinese AI model through its API or self-hosting open weights is not prohibited for most users. However, some Chinese models are subject to US export controls on advanced chips, and the White House has accused specific labs (Moonshot AI, DeepSeek, MiniMax) of industrial-scale distillation of US models. The allegations are not court findings. If your organization has compliance requirements around data sovereignty or supply chain, trace the data path and consult your legal team. (Sources: White House OSTP memo, Anthropic distillation report)
Q: How much does DeepSeek V4 Pro cost?
A: DeepSeek V4 Pro costs $0.435 per million input tokens and $0.87 per million output tokens, with cache-hit input at $0.003625 per million. A 75% launch discount was made permanent on May 22, 2026. These are the lowest prices among frontier-adjacent models as of July 2026. (Source: DeepSeek official pricing, verified via BenchLM)
Q: Can I run Chinese AI models locally?
A: You can run smaller open-weight Chinese models locally — DeepSeek R1 distills (1.5B to 70B), Qwen small tiers, and quantized GLM checkpoints. But the full frontier models (DeepSeek V4 Pro at 1.6T params, Kimi K3 at 2.8T, GLM-5.2 at ~753B) require multi-GPU infrastructure. A 1.5-terabyte BF16 checkpoint does not fit on consumer hardware. Self-hosting a frontier-class Chinese model is a systems engineering project, not a download. (Sources: DeepSeek R1 repository, Hugging Face GLM-5.2)
Q: What is the Kimi K3 distillation accusation?
A: On July 22, 2026, White House OSTP Director Michael Kratsios publicly accused Moonshot AI of distilling Anthropic's Fable model to build Kimi K3. Anthropic had previously (February 2026) accused Moonshot of generating 3.4 million fraudulent Claude exchanges. A cross-entropy forensic analysis by Redwood Research suggested K3 identifies as Claude at elevated rates, though this has innocent explanations (training data contamination, system prompt leakage). The allegations are not court findings, and Moonshot has not publicly responded. (Sources: Business Insider, Anthropic)
Q: Which Chinese AI model is best for coding?
A: GLM-5.2 is the strongest open-weight model for long-horizon coding under a true MIT license, scoring 62.1 on SWE-Bench Pro (vendor-reported, ranked #1 on Artificial Analysis's open-weights Intelligence Index). Kimi K3 scores 81.2 on FrontierSWE (vendor-reported) but at premium pricing ($15/M output). DeepSeek V4 Pro is described by CAISI as strong in software engineering and more cost-efficient than comparable US models on 5 of 7 benchmarks. Test each against your specific coding workload. (Sources: Artificial Analysis, NIST CAISI)
Discussion
0 comments