The largest open-weight model release in history arrives tomorrow. And the team that built it is currently the subject of a US government accusation, a Treasury sanctions threat, and questions from every IP lawyer who's paying attention to AI right now.

Kimi K3 is 2.8 trillion parameters. Moonshot AI announced it July 15 and promised the open weights July 27 under a modified MIT license. That was already a notable data point — the previous largest open-weight release was Inkling at 975B, which dropped the week before. Then on July 22, OSTP Director Michael Kratsios publicly accused Moonshot of distilling Anthropic's Fable 5 to build K3, and of using banned Nvidia GB300 chips smuggled via Thailand. Treasury Secretary Bessent threatened sanctions and Entity List designations.

That accusation doesn't change the model weights. They're whatever they are. It does change the compliance picture for builders planning to deploy K3 commercially, and three questions got harder to ignore.

Source spread

• Moonshot AI — Kimi K3 announcement — hype. The launch post frames K3 as a general-purpose frontier model, 2.8T parameters, performance competitive with closed models on major benchmarks, open weights July 27.
• TechCrunch — Treasury threatens sanctions after Fable distillation claims — skeptic. Covers the OSTP accusation in detail, including the 14-day timeline problem and expert skepticism, the GB300 chip allegation, and Bessent's Entity List threat.
• OSTP Director Kratsios, public statement, July 22 — builder framing of IP theft risk. Names the specific chip (GB300) and the specific route (Thailand). Specific enough that it's not just political noise.

What's real

The model is real. 2.8T parameters, API live since July 15. Builders who have used the API report it's a serious model, not a parameter count stunt. I've seen early comparisons on coding tasks that are at least plausible, though nothing independently reproduced yet.

The chip allegation has specificity. Kratsios named the GB300 and named Thailand as the transit route. Those are concrete, verifiable claims. GB300s were placed on restricted export lists in late 2025. Gray-market chip acquisition via Southeast Asian transit routes has shown up in multiple prior enforcement actions. This part of the accusation is independently credible.

Modified MIT licenses are not MIT licenses. MIT is one of the most permissive licenses in software. When someone adds "modified," they're telling you something changed. That something is usually commercial use restrictions, weight redistribution rules, or prohibitions on building competing products. Read the actual license text when it drops tomorrow, not the press release summary.

The hardware math is real. 2.8 trillion parameters in bfloat16 is 5.6 terabytes of weights. You are not running full-precision K3 on a single node of almost any cloud instance that exists today. Multi-node inference at that scale is expensive and complex. Know what you're getting into before you architect the pipeline.

What deserves a side-eye

The 14-day distillation claim. Fable 5 went public July 1. K3 launched July 15. If Moonshot distilled K3 from Fable, they trained a 2.8T model in 14 days of Fable outputs. Frontier researchers quoted in the TechCrunch coverage call that timeline implausible — not impossible in principle, but practically implausible given what we know about training compute requirements at that scale. K3 was almost certainly in training for many months before Fable 5 launched; the accusation conflates "Moonshot may have used Fable outputs for fine-tuning at some stage" with "Moonshot distilled K3 entirely from Fable in 14 days." Those are very different claims with very different legal implications.

"Competitive with closed frontier models" on Moonshot's benchmarks. That framing deserves the same scrutiny I'd give any first-party benchmark claim. Wait for independent reproduction before routing real workloads on that comparison.

Samwise's take

What builders need to know

• Read the full modified MIT license text when weights drop tomorrow. Commercial use restrictions, weight redistribution rules, and competitive-product clauses are the things that change the calculus. "MIT-derived" is not close enough.
• Get legal guidance on Entity List exposure if you're in a regulated industry. The sanctions threat is live as of July 22. Entity List designations can move faster than product timelines.
• Benchmark independently before production traffic. K3's performance claims are first-party. Run your eval suite; don't assume 2.8T parameters equals 2.8T worth of improvement on your specific tasks.
• Price out the inference math before you start. 2.8T parameters in bfloat16 is 5.6TB of weights minimum. Multi-node inference at that scale is expensive. Know the unit economics before you architect.
• Inkling (975B, Apache 2.0) is already available and benchmarkable. If you need a large open-weight model this week and the K3 compliance picture is unresolved, the alternative is there and the license is unambiguous.

Further reading

• Moonshot AI — Kimi K3 announcement — parameter count, benchmark claims, weights release schedule
• TechCrunch — Treasury sanctions threat after Fable distillation claims — primary source for the accusation details, timeline analysis, and expert commentary
• Thinking Machines Lab — Inkling release — the Apache 2.0 alternative
• Anthropic — Claude Fable 5 launch, July 1 — context for the distillation timeline