# Are Chinese AI Models Really Catching the US Frontier? > Source: > Published: 2026-07-28 00:00:00+00:00 # Are Chinese AI Models Really Catching the US Frontier? DeepSeek, Kimi K3, Qwen, and GLM compared on price, licensing, and capability against US frontier models, with a practical framework for testing them. ## Is China closing the AI model gap? Partially, and unevenly. Chinese labs now field models that compete with US frontier systems on specific tasks like coding, math, and tool use, sometimes at a fraction of the price. But a US government evaluation (CAISI) found DeepSeek V4 Pro, the strongest Chinese model it tested, running about eight months behind the leading US frontier model overall, with the gap shrinking or disappearing on individual benchmarks. That spikiness, strong here, weaker there, is the real story. There’s no single “Chinese model” to measure against a single “US model.” There are several labs pursuing different strategies, and the right comparison depends entirely on the task you’re testing. ## TL;DR **Chinese models aren’t one category**: DeepSeek, Kimi K3, Qwen, GLM, and MiniMax differ wildly in price, licensing, and deployment requirements, so treating “Chinese models” as a single bucket leads to bad decisions.**DeepSeek wins on inference economics**: its API charges roughly 87 cents per million output tokens versus $15 for Kimi K3, making it a strong candidate for high-volume, reviewable work like extraction, classification, and first-pass research.**Cost per token isn’t cost per finished task**: CAISI found DeepSeek ranged from 53% cheaper to 41% more expensive than the US baseline per correctly solved task across seven benchmarks, because cheaper models sometimes need more tokens or more retries to get a right answer.**Open weight means different things for different labs**: GLM 5.2 is MIT licensed but ships a checkpoint around 1.5 terabytes in BF16, Kimi K3 was called open weight before its weights were even downloadable, and MiniMax M3 carries a custom license with revenue thresholds and military-use restrictions.**Mixture of experts explains the price paradox**: DeepSeek V4 Pro has 1.6 trillion total parameters but activates only a fraction per token, which is why a model can be cheap to call through an API and unrealistic to run on personal hardware.**Distillation disputes are already here**: Kimi K3 was publicly accused of distilling from Anthropic’s Claude (“Fable”) family, and Anthropic has said DeepSeek, Moonshot, and MiniMax used fraudulent accounts to pull over 16 million exchanges from its API while evading access restrictions.**For high-stakes or ambiguous judgment calls**, the strongest US frontier systems remain the baseline you should assume you need to beat, given their more mature tool integrations and enterprise controls. ## What are the major Chinese frontier models, and how do they differ? Four names come up constantly: DeepSeek, Kimi (from Moonshot AI), Qwen (Alibaba), and GLM (Z.ai), plus MiniMax as a rising fifth. DeepSeek is optimized for extreme inference economics. DeepSeek V4 Pro runs 1.6 trillion total parameters yet prices output at around 87 cents per million tokens, which is what triggered most of the “China caught up” narrative after the original DeepSeek R1 release. Qwen is less a single model than a broad family, spanning small models you can run locally up to a separate hosted “Max” line whose weights stay closed. That range matters: a small Qwen model and a frontier Qwen model are not the same product, even if they share a name. GLM 5.2, from Z.ai, is MIT licensed and considered particularly strong on long-horizon coding tasks. But the released BF16 checkpoint is enormous, around 1.5 terabytes, which means “open” doesn’t mean “easy to run.” Kimi K3, from Moonshot AI, has 2.8 trillion parameters and a million-token context window, priced as a premium product at roughly $15 per million output tokens. Moonshot has called it open weight, though the downloadable weights lagged behind the model’s release. MiniMax M3 combines coding, computer use, and image and video understanding, with about 427 billion total parameters and only 23 billion active per token. Its license requires attribution, prior written authorization above $20 million in annual revenue, and prohibits military use, a very different arrangement from GLM’s permissive MIT license. The takeaway: these labs share a country of origin and not much else. Price, licensing, deployment burden, and target use case vary by product, not by nationality. ## Why does “open weight” mean so many different things? Because open is not one property, it’s a bundle of separate questions you have to ask individually: Can you call the model through an API? Can you actually download the weights? Does the license permit your specific use, commercial or otherwise? Can your hardware run it at the context length and speed you need? A model can be open on one dimension and closed or impractical on the rest. Kimi K3 was described as open weight before the weights were downloadable. GLM 5.2’s MIT license is genuinely permissive, but the 1.5-terabyte checkpoint size makes self-hosting a serious infrastructure project, not a laptop task. MiniMax M3’s custom license adds legal review to the deployment checklist, especially past the $20 million revenue threshold. Mixture of experts architecture adds another layer of confusion. DeepSeek V4 Pro’s 1.6 trillion total parameters and MiniMax M3’s 23 billion active parameters per token explain why a model can be cheap to run through an optimized API and completely unrealistic to run on personal hardware. Active parameters affect compute per token, but total parameters still drive storage, memory, networking, and serving requirements. Checkpoint size, precision, context length, and serving topology all matter before you can estimate what you actually need to deploy a model yourself. ## Do lower prices actually save money? - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. Not automatically, and this is the part most coverage skips. The CAISI evaluation didn’t just compare sticker prices, it measured cost per correctly solved task across seven benchmarks. The result: DeepSeek ranged from 53% cheaper to 41% more expensive than the US baseline, depending on the task. Cheaper tokens sometimes required longer reasoning traces or more retries to reach a correct answer, and sometimes never got there at all. This mirrors a pattern seen across the industry: a cheap model can become expensive if it produces long reasoning chains, makes unnecessary tool calls, fails late in a workflow, or hands more cleanup to a human. A premium model can end up cheaper overall if it succeeds in one pass. That’s part of why newer frontier releases increasingly compete on token efficiency, not just raw capability. The practical rule: measure cost per accepted result, not cost per token. Test your actual task, your actual harness, and your actual acceptance bar. DeepSeek has also changed its prices since the CAISI evaluation, so any specific percentage is a snapshot, not a permanent verdict. It’s a reminder to test now, not to trust last quarter’s numbers. ## When should you actually use a Chinese model? Break the decision into three separate calls: the task, the model, and the deployment path. Don’t let one choice imply the other two. For high-volume, price-sensitive API work such as extraction, classification, first-pass research, or test generation, where output is reviewable and errors are recoverable, DeepSeek belongs in your test set. Cutting output cost by 15x to 30x changes the economics of running these jobs at scale, and you can afford extra passes to catch mistakes. For local or offline work, on a personal laptop or private server, look at smaller Qwen models or DeepSeek’s distilled models rather than full frontier checkpoints. These are meaningfully smaller systems, not shrunk versions of the flagship model, and they trade capability for privacy, predictability, and offline availability. For long-horizon coding, research, or agent work, test GLM, Kimi, Qwen, and MiniMax directly against US frontier models using the same tools and the same acceptance standard. GLM 5.2 is notably strong at coding, but that doesn’t make it the right choice for every coding task, since a model can write an excellent patch and still fail a repository-level job by picking the wrong tool or missing a test. For sensitive data, decide where the data may go before picking a model. A Chinese model self-hosted on your own infrastructure carries a very different risk profile than the same model accessed through a first-party hosted service based in China. For ambiguous work requiring judgment calls, recovery from unexpected situations, or high-stakes decisions, the strongest US frontier systems remain the standard to beat, given their more mature tool integrations and enterprise controls. ## What’s the distillation controversy about? Distillation itself is completely ordinary. A powerful “teacher” model generates examples, answers, or reasoning traces, and a smaller “student” model trains on those outputs, inheriting some behavior without reproducing the teacher’s full training process. OpenAI sells distillation as a product, and DeepSeek has openly disclosed using R1-generated samples, including 800,000 curated samples, to fine-tune smaller Qwen and Llama-based student models. ## Plans first. *Then code.* Remy writes the spec, manages the build, and ships the app. The controversy starts when a teacher’s outputs were never authorized for that use. Kimi K3 was publicly accused of distilling from Anthropic’s Claude models. Anthropic has said DeepSeek, Moonshot, and MiniMax used roughly 24,000 fraudulent accounts to pull more than 16 million exchanges from its API while evading access restrictions, and a later White House memo described industrial-scale extraction by China-based entities without naming specific labs. The practical implication: a distilled student model is not a compressed copy of its teacher. It inherits useful patterns on the kinds of examples it saw, but loses breadth and reliability elsewhere, which is one reason a small local model and a full frontier model behave very differently even when people casually call both “DeepSeek” or “Kimi.” ## Frequently Asked Questions ### Is DeepSeek actually cheaper than US models for real work? Often, but not always. Token pricing is dramatically lower, but a US government evaluation found DeepSeek’s cost per correctly solved task ranged from 53% cheaper to 41% more expensive than the US baseline across different benchmarks, because cheaper models sometimes need more reasoning steps or retries to reach the same result. ### Can I run Kimi K3 or GLM 5.2 on my own hardware? Not easily. GLM 5.2’s BF16 checkpoint is around 1.5 terabytes, and Kimi K3 has 2.8 trillion parameters. Both require serious enterprise-grade infrastructure, not a personal machine. For local work, smaller Qwen models or DeepSeek’s distilled variants are the realistic option. ### Does open weight mean I can use the model commercially without restrictions? Not necessarily. Licenses vary widely: GLM 5.2 is MIT licensed, while MiniMax M3 requires attribution, prior written authorization above $20 million in annual revenue, and prohibits military use. Always check the specific license before assuming freedom to deploy. ### How far behind US frontier models are Chinese models right now? A CAISI evaluation estimated DeepSeek V4 Pro, the most capable Chinese model tested at the time, was about eight months behind the leading US frontier model overall, though the gap narrowed or vanished on specific tasks like coding and math. ### What’s the safest way to start testing Chinese models? Separate your decisions: pick a bounded, reviewable task first (like classification or extraction), test a cost-focused model like DeepSeek against your current tool using identical prompts and acceptance criteria, and measure cost per accepted result rather than cost per token before expanding to more ambiguous or high-stakes work.