When I read the claim that a Chinese AI lab had trained a model with 2 to 3 trillion parameters—positioning it as a direct challenge to Anthropic’s Claude 3.5—I didn’t feel excitement. I felt a familiar chill.
That chill is the same one I felt in 2020 when a DeFi protocol I was auditing claimed to have 'unparalleled liquidity depth' right before we found a reentrancy vulnerability in their flash loan module. It’s the chill of a narrative built on a single, unverifiable metric.
We built trust in the chaos, not despite it. In crypto, we learned the hard way that total value locked means nothing without audited contracts. In AI, total parameter count means nothing without benchmark results, training transparency, and safety disclosures.
So let’s unpack the Kimi K3 story—not as a blockchain analyst, but as someone who has spent years watching industries conflate scale with substance.
The Context: A Single Number, A Big Claim
The report states that Moonshot AI’s Kimi K3 model boasts 2 to 3 trillion parameters. That would make it larger than GPT-4’s estimated 1.8 trillion and Claude 3.5’s ~2 trillion. On the surface, it’s a headline: “China’s largest model challenges Anthropic.”
But I’ve learned that in technical systems, the surface is often a carefully polished lie. My 2017 experience founding ChainBridge taught me that the real value isn’t in the size of the smart contract library, but in how the community uses it. My 2026 work on the Human-in-the-Loop standard for AI governance taught me that algorithmic outputs without human ethics are dangerous, regardless of how many neural connections they have.
Here’s what we don’t know about Kimi K3: No MMLU scores, no GSM8K results, no HumanEval pass rates, no context length ceiling, no training efficiency numbers, no safety audit results. Nothing but a parameter count.
The Core: Why Parameters Are the New TPS
In blockchain, we once obsessed over transactions per second as the ultimate metric. Projects claimed 100,000 TPS without showing how that throughput handled real-world congestion, latency, or security. We learned that TPS without decentralization is just a centralized database.
Parameters are the new TPS. A model with 2–3 trillion parameters is almost certainly a Mixture-of-Experts architecture. The total parameter count includes the router and all experts, but the activated parameters per inference are likely 200–300 billion—equivalent to Claude 3.5 and GPT-4. The headline number is designed for press releases, not for engineers.
I’ve been in enough audits to know that when a team emphasizes one metric over all others, they are hiding something else. In 2020, the OpenYield team talked about their “innovative yield curve” while I was finding the vulnerability in their flash loan logic. Here, Moonshot AI emphasizes raw size while remaining silent on every dimension that actually matters for users: reasoning, safety, cost, and accessibility.
Furthermore, scaling laws have diminishing returns. Training a 3-trillion-parameter model requires roughly 210 trillion tokens under the Chinchilla optimal regime. That’s far beyond the size of any public dataset. Either Moonshot AI used synthetic data, scraped private conversations, or cut corners on quality. In crypto, we call that a liquidity problem—you can inflate the numbers, but eventually the market finds out.
The Contrarian Angle: Maybe Parameter Count Still Matters for Fundraising
Here’s the uncomfortable truth I’ve learned from 2022's bear market: hype can raise capital, even when the product doesn’t deliver. During the FTX collapse, I saw projects raise millions on the back of “institutional-grade security” without a single external audit. The narrative is a funding vehicle.
Moonshot AI has reportedly raised around $1 billion. If the K3 announcement attracts a new funding round at a $5–6 billion valuation, the bet pays off—even if the model never beats Claude 3.5 on any benchmark. The investors are buying a story, not a product.
But as an educator, I see that story as a trap. Education is the antidote to exploitation. When we teach developers to judge models by parameter count, we train them to make the same mistake we made with TPS. We create a market where real innovation is overshadowed by marketing.
The Infrastructure Contradiction: Where Are the Chips?
Training a 2–3 trillion parameter model requires 10,000 to 20,000 H100 GPUs running for months. The US export controls on H100 and A800 have made such a cluster nearly impossible to assemble in China legally. The only path is using domestically produced chips like Huawei’s Ascend 910B, which have lower real-world utilization.
If Moonshot AI managed to train K3 on domestic chips, that would be a significant engineering achievement—and they would have every incentive to brag about it. Their silence on the training infrastructure is deafening. It suggests either the cluster relied on smuggled H100s (risking future supply and national regulation) or the model wasn’t fully trained as claimed.
From my experience in 2024 bridging Wall Street and Web3, I know that institutional partners demand verifiable infrastructure details. The lack of transparency here will limit K3’s enterprise adoption, regardless of parameter count.
The Security Void: Code Is Law, But Humans Are the Protocol
Anthropic has built its brand on safety. They published their Constitutional AI paper, released red teaming results, and integrated with AWS’s security suite. Moonshot AI, by contrast, has released no safety information for K3. No harmlessness benchmarks, no jailbreak resistance tests, no data privacy disclosure.
In the 2022 bear market, I saw what happens when projects ignore safety—they become platforms for scams and user loss. The same will happen with AI models that lack alignment. A model with 3 trillion parameters that can be easily jailbroken is more dangerous than a smaller, safer model.
The Chinese regulatory environment requires AI models to pass safety assessments for public release. But compliance with local censorship doesn’t replace global safety standards. Trust is earned in drops, lost in buckets.
Takeaway: The Future Belongs to Those Who Teach Together
When I launched The Anchor Project after FTX, I learned that real value comes not from the size of your platform, but from the trust you earn by being transparent, ethical, and present. The same applies to AI.
Moonshot AI’s K3 might be a genuinely good model. But by hiding behind a single parameter count and refusing to publish benchmarks, safety audits, or training details, they are repeating the exact mistakes that caused the crypto collapse of 2022. They are choosing narrative over substance.
Hold through the noise, build through the silence. The noise around K3 will fade. The silence about its real performance will echo. And in that echo, I hear a lesson we already learned: trust must be earned, not claimed.
I’ll wait for the independent benchmarks. Until then, I’ll keep teaching my students that the most important number in any system isn’t the count of parameters or the TPS, but the probability that the system does what it claims to do.