What this is
321B parameters puts it in GPT-4 territory. This week, open-source community Project Maya got a Chinese model of that size—Zhipu AI's GLM-5.3-Flash—running on a single consumer RTX 5090. The trick is tiered scheduling: under the model's MoE architecture, the network is split into multiple smaller expert groups that are called on demand. The frequently used experts live in VRAM, less-used ones in system RAM, and rarely used ones on SSD. They get shuffled in as needed during conversation. Measured throughput: 30+ tokens/sec on a single card, 118 tokens/sec on four 4090s combined. The license is MIT (free commercial use allowed), and quantization (compressing the model to save VRAM) retains 97.7–99.2% of original accuracy. Equally important: API compatibility with both OpenAI and Anthropic—apps that previously called GPT can switch to a local endpoint without rewriting code.
Industry view
The bull case: For the first time, a Chinese model stands shoulder-to-shoulder with Meta's LLaMA family on the "open-source weights + community inference engine" path. Open source, strong Chinese capability, and data staying on-prem—this is a combination enterprise legal and procurement teams both understand.
Worth watching on the downside: 1) The 97.7% accuracy figure is developer self-reported, with no third-party benchmark behind it. 2) On-demand shuffling causes latency jitter, and long chained tasks may slow down. 3) Single-card solutions remain hostage to GPU pricing—an RTX 5090 at 16,000 yuan (~$2,200) is not typical household spending. 4) Zhipu faces heavy commercialization pressure in 2025; whether it can keep investing in the community is an open bet.
Our editorial take: The real meaning here is that the price sheet for "private deployment" is starting to crack. What to watch in the second half of the year is whether enterprise integrators have the nerve to cut their solution budgets in half.
Impact on regular people
For enterprise IT: The default eight-card H100 (data-center grade GPU) configuration is now back on the negotiating table—procurement budgets can be compressed by at least an order of magnitude.
For working professionals: Running a large model locally to draft internal contracts and industry reports means sensitive data can, for the first time, avoid touching any cloud at all.
For the consumer market: Short-term boost for high-end consumer GPUs; the real winners are the integrators packaging "enterprise private AI solutions."