Chinese Open-Weight Labs Face Revenue Pressure
Z.ai and MiniMax reported steep 2025 losses despite growing revenue, highlighting the cost of building competitive open-weight models. Z.ai disclosed RMB 724.3 million in revenue and a RMB 3.18 billion adjusted net loss; MiniMax reported $79.0 million in revenue and a $250.9 million adjusted net loss. Open weights reduce distribution barriers, but they do not remove training, inference, or operating costs.
Z.ai and MiniMax reported steep 2025 losses even as Chinese open-weight models gained developer attention. The disclosures give concrete numbers to a July 21 Business Insider analysis about the challenge of building a durable business around model weights that users can download or run through competing infrastructure providers.
What the filings show
Knowledge Atlas Technology, the listed company behind Z.ai and its GLM models, reported RMB 724.3 million in 2025 revenue. Its annual report recorded a RMB 4.72 billion loss for the year and a RMB 3.18 billion adjusted net loss, with the company attributing the wider reported loss mainly to higher research-and-development investment.
MiniMax reported $79.0 million in 2025 revenue. Its annual report recorded a $1.87 billion loss for the year, including a $1.59 billion fair-value loss on financial liabilities. After excluding that and other specified items, MiniMax reported a $250.9 million adjusted net loss. That distinction matters: the roughly $250 million figure cited in Business Insider is an adjusted measure, not the company's full reported loss.
Business Insider also reported that Z.ai shares had fallen more than 40% over the preceding month and MiniMax shares more than 50%. Those market moves do not establish that open-weight distribution caused the losses, but the filings show that revenue remained far below the companies' spending and accounting losses in 2025.
Why open weights do not create software-like costs
Publishing model weights can make copying and adapting a model easier, but producing and serving the model still consumes scarce resources. Training requires chips, data, engineering, and research talent. Inference requires accelerator capacity, power, networking, and operations for every production workload.
TechCrunch separately reported that neither open nor proprietary AI companies have settled on a reliable business model. Open-weight systems can pressure premium API pricing and broaden adoption, while their developers still need revenue from hosted inference, enterprise deployments, support, applications, or other services. The filings do not prove that every open-weight strategy is uneconomic; they show the scale of the commercialization gap at two prominent independent labs.
What practitioners should compare
Axios reported that Chinese models held the top five positions by weekly token usage on OpenRouter in mid-July, and that companies were routing routine work toward cheaper models while reserving premium systems for harder tasks. That usage signal is relevant, but token volume is not the same as profit.
For technical buyers, the useful comparison is total workload economics: model quality, tokens per second, latency, accelerator utilization, data movement, reliability, observability, and the engineering required to operate the serving stack. Open weights can improve control and portability, but the cost advantage depends on workload scale and execution rather than the absence of a software license alone.
Key Points
- 1Z.ai reported RMB 724.3 million in 2025 revenue, a RMB 4.72 billion loss for the year, and a RMB 3.18 billion adjusted net loss.
- 2MiniMax—not Baichuan—reported $79.0 million in revenue and a $250.9 million adjusted net loss; its full reported loss was $1.87 billion after a large fair-value charge.
- 3Open weights can improve control and portability, but practitioners still need to compare total serving cost, performance, reliability, and operating effort.
Scoring Rationale
Official 2025 filings from Z.ai and MiniMax document a large gap between revenue and adjusted losses at two prominent Chinese model developers. The figures matter to practitioners evaluating open-weight adoption and self-hosting economics, while the article carefully avoids treating two companies as proof that every open-weight model strategy is unprofitable.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
