Crypto Briefing broke the news of Inkling, a 975B parameter open-source model “built for fine-tuning.” The bubble isn’t the story; the story is the story selling it.
Friction reveals the fault lines no one else sees. Here, the fault line is the source itself: Crypto Briefing is a media outlet deeply embedded in the Web3 ecosystem, not a peer-reviewed AI journal or a technical blog from an established lab. When an unknown entity claims to have trained a model larger than Meta’s Llama 3 405B and makes it open-source, the immediate question is not “Is it good?” but “What are they really selling?”
Context: The Familiar Playbook The crypto industry has a long history of borrowing AI hype to prop up token valuations. During the 2021 NFT boom, countless projects slapped “AI-generated” onto their roadmaps to attract capital. In 2024-2025, the convergence narrative became more sophisticated: decentralized compute networks, AI agent tokens, and models offered as collateral for lending protocols. Thinking Machines, the entity behind Inkling, fits this pattern perfectly. No public team credentials, no whitepaper detailing training methodology, no benchmarks. Just a press release with a massive number and a button labeled “open source.”
Core: The Technical Red Flags Let’s be honest: a 975B parameter model is not trivial to train. Even with today’s most efficient hardware (H100 clusters), the compute required exceeds $10 million in cloud costs. Yet Thinking Machines provides zero details on their infrastructure, training duration, or data composition. They claim the model is “built for fine-tuning,” which is a convenient way to hide a weak base model. Compare this to Llama 3 405B: Meta published extensive evaluations, safety red-teaming results, and even a technical report. The contrast is stark.
Moreover, the parameter count itself is suspicious. 975B is an odd number—most models round to 1T or stick with powers of two. It suggests the number might be inflated by embedding layers or vocabulary size rather than actual reasoning capacity. In my experience auditing DeFi contracts, I’ve seen similar tricks: a protocol claims $1B TVL but includes locked tokens it controls itself. Parameter counts can be gamed the same way.
The market doesn’t care about your feelings. But it should care about the asymmetry of information. Right now, the only information available is the narrative propagated by Crypto Briefing. Any serious investor or developer should demand independent verification before even considering deployment. This is not skepticism—it’s survival.
Contrarian: Why This Might Actually Be Good for Crypto Now for the twist. Despite my skepticism, Inkling could serve a real purpose: stress-testing the decentralized GPU market. If Thinking Machines truly open-sources the weights and enables permissionless fine-tuning, it will force platforms like Akash, Golem, and io.net to prove they can handle models of this scale. The demand for decentralized compute has always been limited by the lack of large models to run. Inkling, even if mediocre, could be the catalyst that pushes DePIN infrastructure to maturity.
Furthermore, the “Crypto Briefing” connection is a canary in the coal mine. If this is a prelude to a token launch, the regulatory scrutiny will be intense. The SEC has already fined projects for unregistered securities tied to AI tokens. A failed token launch could set precedent that discourages future “AI+blockchain” scams, cleaning up the space for legitimate builders.
Takeaway: Watch the Community, Not the Headline In the next 30 days, follow the Hugging Face page. If the model is actually uploaded and the community begins to replicate benchmarks, then we have a real asset. If instead we see a token airdrop announcement first, run. The bubble isn’t actually the model—it’s the story being sold around it. Friction reveals the fault lines: the gap between a press release and a working product is where the risk lives.
Based on my background dissecting DAO governance failures in 2020 and NFT rug pulls in 2021, I recognize this pattern instantly. The market will eventually price in the truth—but only if we demand it now. Don’t let the allure of a 975B parameter number blind you to the absence of substance. The real innovation isn’t the parameter count; it’s the infrastructure that can actually run it without a centralized master key. That’s the story worth watching.