Two Chinese model labs reported first-half 2026 results. Zhipu posted RMB 954 million in revenue and an adjusted net loss of RMB 1.964 billion. MiniMax posted $117 million in revenue and a $293 million loss.
Key takeaways
- Market cap on 4 September: Zhipu around HKD 500 billion, MiniMax around HKD 126 billion
- Zhipu: platform and API account for RMB 825 million, or 86.5% of revenue
- MiniMax: open-platform revenue grew 703.1% year on year, to $73.93 million
- Token consumption: over 40x since the start of the year at Zhipu, 20x from January to July at MiniMax
- Gross margin fell at both: Zhipu to 26.4% from 50%, MiniMax to 17.9% from 25.4%
Two labs side by side
| Metric | Zhipu | MiniMax |
|---|---|---|
| H1 2026 revenue | RMB 954M | $117M |
| Adjusted net loss | RMB 1.964B | $293M |
| Market cap (4 September) | approx. HKD 500B | approx. HKD 126B |
| Platform and API share of revenue | 86.5% | 63.4% |
| Gross margin (year earlier) | 26.4% (50%) | 17.9% (25.4%) |
| Token consumption growth | over 40x since year start | 20x from January to July |
The two companies report in different currencies, so proportions and momentum are comparable while absolute values are not. Both market caps are given in Hong Kong dollars as of 4 September 2026.
Zhipu: the value has moved entirely into the API
The revenue mix at Z.ai inverted within a year. On-premise deployments now bring in only RMB 129 million, or 13.5% of the total. The remaining RMB 825 million comes from the open platform and API. Call volume for tokens grew more than 40-fold since the start of the year, and Coding Plan calls alone rose more than 23-fold.
Notably, the average API price went up by roughly 101%, so the growth is not coming from discounts. Even so, gross margin?Gross margin: The share of revenue left after direct costs of delivering the service. For AI models the largest line item is inference — the compute spent on every single request. fell from 50% to 26.4%.
MiniMax: faster growth, from a lower base
MiniMax grew revenue 283.1% year on year, and its open-platform segment 703.1%, to $73.93 million. The B2B share of ARR?ARR: Annual Recurring Revenue — the value of active subscriptions and contracts projected across a full year, usually extrapolated from the current month. jumped from around 30% to roughly 80% in a year.
The company runs two tracks at once, though. Its public model repository makes that visible.
| Model | Modality | Parameters |
|---|---|---|
| MiniMax-M3 | image + text → text | 427B |
| MiniMax-M3-MXFP8 | image + text → text | 440B |
| MiniMax-M2.7 | text generation | 229B |
| MiniMax-H3 | image + text → video | 33B |
| MiniMax-Music3 | text → audio | 2B |
The video and audio line builds reach — H3 is close to 5 million downloads — but there is no sign it builds revenue.
Token sales move to the retail shelf
This is the point at which model capability gets a common price tag. The most telling signal of the half-year is therefore not in the accounts but in the sales channel.
Why it matters
A fourfold valuation gap on similar losses says more about narrative than about the income statement. The market rewards a company with one legible axis — chat, coding, agent — and discounts the one pulling text and video in parallel.
The gross-margin decline at both, meanwhile, shows that sheer token volume does not improve profitability. Inference cost grows alongside demand.
What's next
- Zhipu's gross margin fell to 26.4% despite a 101% API price increase — the next half-year will show whether the increase is enough to reverse the trend
- MiniMax has to settle whether the open H3 video model stays a marketing instrument or gets a revenue model of its own
- Listing on Tmall creates directly comparable pricing across vendors, which at 40x volume growth could trigger a price war
Sources
- TMTPost — 确定的智谱，不确定的MiniMax
- TMTPost — 大模型的Token生意，终究没绕过天猫
- Hugging Face — MiniMaxAI model repository
