A Reddit user pitted two Chinese open-source models against each other on two NVIDIA DGX Sparks (~$8,500 each) and surfaced a number: Zhipu's GLM-5.3 Flash beat DeepSeek V4 Flash on a code-writing test 97% to 94.5%. Open-source small models on "small machines" have truly entered a trade-off war — GLM runs 30% slower, and its context window is only a quarter of DeepSeek's.
What this is
HumanEval is the industry's standard test for "can AI write code." With deep-thinking mode enabled, GLM-5.3 Flash hit a 97% pass rate; DeepSeek V4 Flash scored 94.5%. GLM finished the run in 20 minutes 52 seconds; DeepSeek needed 38 minutes 16 seconds.
The cost comes on the other side: GLM generates about 50 tokens per second; DeepSeek about 70. GLM's context window (how much text AI can process at once) is 256,000 characters/tokens; DeepSeek's is 1 million. In plain terms: GLM writes code more accurately; DeepSeek handles long text better.
Industry view
Official rhetoric will hype domestic open-source models. But we flag three cooling points:
This is a Reddit individual's test, not an institutional benchmark — the author themselves notes at the end: "my own experience matters more than any benchmark." HumanEval only measures code writing, a narrow metric — you can't derive "comprehensive surpassing" from it. A 2.5 percentage-point gap isn't statistically meaningful, and both models used quantization (a technique that stores the model in less space, at the cost of some precision). Declaring "complete reversal" is not rigorous.
Looked at the other way: three years ago, hitting 97% inside 256GB of VRAM was science fiction. The pace at which open-source catches closed-source is genuinely accelerating.
Impact on regular people
For enterprise IT: "running AI on the intranet" used to be a tech-nerd pursuit; now there's an off-the-shelf path. Finance, healthcare, and government — industries sensitive to data compliance — should ask at procurement time: can we run this ourselves?
For individual professionals: the daily office routine won't change in the short term. But cheaper models mean subscription-based AI tools will keep dropping prices — expect a wave of promotions before year-end.
For the consumer market: phone and PC vendors will market "on-device large models" more aggressively over the next year — models that run directly on the device, no uploading required. Chinese AI on phones is shifting from "barely usable" to "actually getting work done."