Open-weight AI just proved it can print real money. China’s Moonshot AI, the lab behind the Kimi assistant, is now targeting $2 billion in annualized revenue by year’s end. That’s double its reported run rate from August. According to TechCrunch AI, which cited a Friday Bloomberg report, the goal rides on the breakout success of the company’s K3 model released this summer.
Here’s what stands out: Moonshot is making this play with freely available model weights. That’s the hard road to revenue, and they’re walking it anyway.
Situation Report
- TARGET: $2B annualized revenue by end of 2026, up from a run rate half that size in August.
- DRIVER: The K3 model. TechCrunch AI reports that OpenRouter data shows as many as 300 billion tokens generated daily by K3 models on the platform, even with usage slipping a bit in recent months.
- SCALE CHECK: Moonshot is still a minnow next to the frontier labs. Recent figures put OpenAI near $40 billion and Anthropic around $65 billion.
- MARGIN PROBLEM: Because the weights are open, Moonshot earns far thinner margins than its closed-weight rivals. Anyone can run the model themselves.
Why This Matters
The open-weight versus closed-weight debate has been mostly philosophical. Open models were seen as community goodwill, research tools, loss leaders. Moonshot is turning that assumption on its head.
A $2 billion run rate says there’s a real business in open-weight AI, even if it never matches the economics of a locked-down frontier model. For practitioners, that’s the headline. The open ecosystem you build on isn’t a charity project that might vanish when the funding dries up. It can stand on commercial legs.
That changes planning. If you’re betting your stack on open weights, a vendor with billions in revenue is a safer bet than one burning cash with no path to sustainability.
Threat In The Mix
Now the complication. Moonshot’s training methods are under fire, and the accusation is serious.
Earlier this week, per TechCrunch AI, Anthropic accused Moonshot of a long-running model distillation campaign. The claim: nearly 300,000 requests from Kimi were routed directly to Claude Opus, effectively serving Opus responses in place of Kimi’s own model output. In total, Anthropic alleges more than 23 million responses were harvested from its models to train Moonshot’s systems.
If that holds up, it isn’t just a bad look. Distillation of this kind typically violates the terms of service of the model being copied, and it raises the question of how much of K3’s quality is actually Moonshot’s own work versus borrowed intelligence from a competitor.
Model distillation, briefly explained: you feed prompts to a stronger model, collect its answers, and train your own smaller or cheaper model to imitate them. It’s a well-known shortcut to capability. Done with permission, it’s routine. Done by quietly piping your users’ queries to a rival’s paid model, it’s a different story.
What To Watch
- ENFORCEMENT: Whether Anthropic moves beyond accusation to legal action or API-level blocks. That could slow Moonshot’s model pipeline.
- REVENUE PROOF: Whether the $2 billion target is real demand or an aggressive projection. Token volume is promising, but a dip in K3 usage is worth tracking.
- REGULATORY ANGLE: How US and Chinese authorities treat cross-border distillation disputes, since this pits a Chinese lab against an American one.
- COPYCATS: Other open-weight labs watching Moonshot’s numbers and deciding the commercial model is worth chasing.
BOTTOM-LINE ASSESSMENT: Moonshot is showing that open-weight AI can generate serious revenue, not just research credibility. That’s genuinely new. But the distillation allegations hang over the whole story, and they could undercut both the company’s reputation and its growth if Anthropic pushes back hard.
This is a rare case where the business milestone and the controversy landed in the same week. Keep both in view. Full details are available at the original TechCrunch AI report.