Chinese vs Western LLM Latency: What We Measured

In our 2026-07-10 snapshot, DeepSeek V4-Flash answered a short prompt on its official endpoint in about 0.7 seconds, while the same class of Chinese models routed through a US aggregator took 2.8 to 7.6 seconds, so the endpoint, not the country, sets latency. That single finding reframes the whole "are Chinese models slow" question: what we timed was mostly the path a request travels, not the nationality of the lab that trained the weights.

This page is our strongest first-hand latency read, and it is deliberately narrow. It is a single short-prompt snapshot, not a throughput benchmark, and it does not time Western endpoints at all. We say where each number came from and where it stops.

Chinese vs western llm latency (verified 2026-07)

Time to first token is a latency metric that measures the delay between sending a request and receiving the first streamed output token. In our short-prompt snapshot on 2026-07-10, DeepSeek V4-Flash on its official endpoint returned that first token in roughly 0.7 seconds, the fastest result we recorded.

The routed Chinese models are a different story, because they did not run on their vendors' own endpoints. We reached Qwen, MiniMax, DeepSeek-V3.2, and GLM-5 through OpenRouter, a US-based model aggregator, which adds a network hop and a host-selection step before the model even starts. Here is the snapshot.

According to DeepSeek API Docs, the deepseek-chat alias is served by the V4-Flash model class, the tier we timed at roughly 0.7 seconds on the official endpoint. DeepSeek operates api.deepseek.com as its own native endpoint, which is why that path was both the fastest and the most reproducible in our runs.

What we measured, first-hand (2026-07-10)

The 0.7-second figure is not a vendor claim. We measured it directly on api.deepseek.com in a cold-start run, and it matches the sub-second first-token behavior we have seen across repeated DeepSeek calls. That number is the anchor for this whole page, and it is the one reading here that came from a model's native endpoint.

Routing overhead is the added round-trip time introduced when a request passes through a third-party aggregator instead of a model's native endpoint. Our routed Chinese models landed between 2.87 and 7.62 seconds on the same short prompt, a spread that reflects the aggregator hop, the host it selected, and each model's own compute time.

Read the routed numbers as an upper bound, not a verdict. A native endpoint is a vendor-operated API that serves a model directly without a reseller layer. Because DeepSeek runs api.deepseek.com itself, our fastest and most reproducible latency came from that path, which is why we treat the routed figures as an upper bound rather than a fair read of native Chinese-model speed.

The gap between DeepSeek's 0.7 seconds native and the routed models' 2.8-to-7.6-second band is therefore a routing-and-host story first. Some of the routed spread is real model difference, a larger model that thinks longer will stream its first token later, but the floor moved up the moment we left the native endpoint.

We did not time Western endpoints here

We did not benchmark Western endpoints for latency in this run. Our GPT-4o and Claude Sonnet 4.6 tests measured price only, billing $0.000795 and $0.001344 on the same prompt, so any speed comparison against them here would rest on vendor claims, not our own timing, and we do not make one.

Where a native Chinese endpoint exists, region is a real variable. According to Alibaba Cloud, Qwen models are served through Model Studio with endpoints in more than one region, so a caller's native latency depends on which regional endpoint they select and how far they sit from it. Alibaba Cloud develops the Qwen family, and its documented multi-region hosting is the reason a native Qwen call from Singapore need not resemble our US-routed 2.87-second reading.

The reason so many of these models are reachable through aggregators at all is their open-weight distribution. According to Stanford HAI, Chinese open-weight models accounted for a majority of tokens consumed on one large routing marketplace by 2026, which is why third-party hosts, and the routing overhead they add, are so common on the China side.

FAQ

Why is a routed Chinese model slower than DeepSeek's native endpoint? Because a routed request travels through a US aggregator that adds a network hop and picks a host before the model starts working. In our 2026-07-10 snapshot the routed Chinese models took 2.87 to 7.62 seconds, versus about 0.7 seconds for DeepSeek V4-Flash called directly on api.deepseek.com. The routing path, not the model's nationality, explains most of that gap.

Does region matter for Chinese LLM latency? Yes, when you use a native endpoint. Alibaba Cloud documents serving Qwen from more than one region, so distance to the chosen regional endpoint changes your round-trip time. Our routed figures do not capture this, because they went through a single US aggregator rather than a nearby native endpoint, so treat them as an upper bound.

Did you measure GPT-4o or Claude latency in this test? No. We tested GPT-4o and Claude Sonnet 4.6 for price only, billing $0.000795 and $0.001344 on the same prompt on 2026-07-10. We did not time their first-token or round-trip latency, so this page makes no speed claim about Western endpoints and defers to their own vendor documentation for that.

Is this a throughput benchmark? No. It is a single short-prompt snapshot captured on 2026-07-10, not a sustained-load or tokens-per-second benchmark. Long generations, concurrency, and streaming behavior can all shift the numbers, so use these figures to compare paths and endpoints, not to size a production workload.

Will native endpoints for Qwen, GLM, and MiniMax be faster than the routed numbers? Very likely, since they skip the aggregator hop, but we have not yet measured them natively. Those rows are flagged needs-native-reverify for exactly that reason, and we will update them once we run each vendor's own endpoint rather than a US-routed host.

This is part of our Chinese vs Western LLM comparison hub. For the price side of the same models, see our DeepSeek API pricing breakdown.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem latency and pricing. Last verified: 2026-07-10.