Kai ChenChinese foundation-model lab MiniMax listed in Hong Kong in January 2026. Its first annual report showed $79 million in 2025 revenue against a $1.87 billion net loss. On July 9, Bloomberg reported that the lab was seeking roughly $2 billion more through shares and zero-coupon convertible bonds due in 2027. Six months into its public life, MiniMax was already asking the market to fund its next stage.
Key takeaways
- MiniMax listed in Hong Kong in January 2026.
- MiniMax reported $79 million in revenue for 2025.
- MiniMax reported a $1.87 billion net loss for 2025.
- The proposed follow-on package included 35.6 million shares priced near $34 and about $830 million in zero-coupon convertible bonds due in 2027.
- MiniMax quoted M3 input pricing at $0.12 per million tokens, versus $5 per million for Claude Opus 4.7.
MiniMax returned to the capital markets before its first year as a listed company was over. Investors now have to judge whether recurring revenue from agent workflows can outrun the continuing costs of serving, improving, and distributing its models. Selling shares and bonds buys MiniMax time, but each transaction puts model pricing, enterprise adoption, and operating efficiency back before the market.
A listing converts runway into a financing schedule
MiniMax entered its January 2026 IPO after reporting $30.5 million in 2024 revenue and a $465.2 million loss. Its first post-listing annual report put the following year’s figures much farther apart.
Investors cannot treat the entire loss as an inference bill; the filing does not attribute every dollar to serving models. But MiniMax still has to pay for model development before demand arrives and for capacity when customers use the models. Revenue growth had not yet produced a self-funding operation.
On May 30, Bloomberg reported that MiniMax had begun preparations for a mainland Chinese listing and said annual recurring revenue had reached $300 million. ARR describes a current annualized pace rather than the prior year’s recognized revenue. Investors still need retention and gross-margin data before comparing that figure with the 2025 loss.
MiniMax then sought to sell 35.6 million shares priced near $34 and about $830 million of zero-coupon convertible bonds due in 2027. The report described a planned transaction, not completed financing. The share sale would place long-duration risk with equity holders, while the bond would put another conversion or repayment decision on the calendar the following year.
By July, MiniMax had opened or prepared access to three forms of capital: Hong Kong equity, a mainland listing, and a combined share-and-convertible sale. Those moves do not prove distress. They show how closely the company’s model-development schedule now depends on continued market access.
Agents turn cheap tokens into a harder margin test
MiniMax priced M2.5 at $0.30 per million input tokens and $1.20 per million output tokens, presenting the model as “intelligence too cheap to meter.” M3 pushed the input quote to $0.12 per million tokens, compared with $5 for Claude Opus 4.7. MiniMax positioned the open-weights model for coding and agentic work with a one-million-token context window.
A developer may pay little for the first prompt while spending far more to finish the task. A long-running agent can invoke a model repeatedly, generate substantial output, call tools, retry failed steps, and carry context forward. MiniMax therefore needs task-level accounting that captures every call and retry rather than the advertised price of one million tokens.
Investor documents reviewed by The Wall Street Journal in April 2026 showed OpenAI and Anthropic reporting inference costs above half of revenue. On January 22, The Information reported that Anthropic had cut its projected 2025 gross margin for enterprise and developer sales to 40% from 50% after inference costs ran 23% above its estimate. Investors cannot transfer those ratios directly to MiniMax, but the figures show how serving costs can compress margins even as revenue rises.
Providers can trade some latency for throughput when a human is not waiting for every response. They can batch requests, route tasks among models, and place workloads where capacity is cheaper. For agent products, those operating choices directly affect the price of completed work.
A team evaluating M3 will budget the entire loop: model calls, retries, tools, and context. It will compare that bill with the value of completing a pull request, resolving a support case, or executing a governed workflow. MiniMax’s investors need the company-side version of the same measure—its AI cost per useful task.
Open weights make hosting the product
MiniMax open-sourced M1 in 2025 with support for one million input tokens and a focus on complex productivity tasks. In 2026, it released H3 for 2K video clips of up to 15 seconds with native stereo sound and said it would release the weights within days.
Developers who receive the weights can inspect, adapt, and deploy the technology without using MiniMax’s proprietary API. They can also choose another host. That route gives MiniMax wider distribution but less control over where inference revenue lands.
Moonshot AI illustrated the same substitution when it released Kimi K3’s weights under a license permitting users to download, modify, and host the model. Both labs accepted third-party hosting as the price of reaching more developers.
A developer who can move the weights needs a reason to keep paying MiniMax. The company must make its hosted service valuable through reliability, workflow integration, enterprise controls, and ongoing model operations—services a weight file cannot provide.
Share prices change MiniMax’s dilution math
After JPMorgan preferred Zhipu AI to MiniMax and raised its price target, Zhipu closed the session up 33%, while MiniMax gained 7%. One trading day cannot settle either company’s economics. It can immediately change financing capacity: at a higher share price, either lab can raise a fixed sum with fewer new shares.
MiniMax’s proposed package made that sensitivity concrete. It paired a sale of 35.6 million shares near $34 with about $830 million in convertibles. A weaker stock price would reduce proceeds from a fixed share count or force MiniMax to sell more shares for the same cash. The bond postpones part of the financing decision until 2027 rather than eliminating it.
Durable revenue begins where the model disappears into work
MiniMax designed M3 around coding, agent capabilities, native multimodal input, and long context. Those features make the clearest commercial sense inside bounded workflows where a buyer can specify permissions, data access, expected outputs, and review.
Apple showed how agents enter those workflows when it integrated Anthropic’s Claude Agent and OpenAI’s Codex into Xcode 26.3 and added MCP support. Anthropic reported that software engineering generated about half of its agent tool calls. Repositories constrain the task, tests expose failures, and review creates an accountability checkpoint.
In November 2024, Microsoft said more than 100,000 organizations were making or editing agents in Copilot Studio and that the count had doubled in one quarter. The figure documents experimentation at scale, not production retention. Enterprises still need controls around permissions, consequential actions, and human review before those experiments become dependable deployments.
Frequently asked questions
Did MiniMax complete the roughly $2 billion follow-on financing?
The evidence describes a planned or sought transaction, not a completed raise. It does not establish final proceeds, pricing, investor participation, or closing date.
What conversion terms apply to MiniMax’s proposed 2027 convertible bonds?
The piece identifies roughly $830 million of zero-coupon convertibles due in 2027, but gives no conversion price, conversion ratio, or other conditions that would determine potential dilution.
What share of MiniMax’s reported $300 million ARR comes from agents or enterprise workflows?
The evidence does not break ARR out by product, customer type, workflow, retention, or geography. It therefore cannot show how much of that annualized pace is durable agent revenue.
Will MiniMax complete a mainland Chinese listing?
MiniMax had begun preparations for a mainland listing as of the May 30 Bloomberg report. The evidence provides no filing outcome, exchange, timetable, or confirmation that the listing will occur.
MiniMax revenue and net loss
| Year | Revenue | Net loss |
|---|---|---|
| 2024 | $30.5M | $465.2M |
| 2025 | $79M | $1.87B |
Six months after its IPO, MiniMax asked the market for roughly $2 billion more. The proposed financing would buy time to turn low token prices into retained workflow revenue. Developers can move open weights among hosts, while every completed task consumes compute. The Hong Kong listing gave MiniMax a live share price; governed, repeated work must give that price something durable to measure.