Open Weights, Closed Governance: Inkling-Small and the Next Decentralization Battle

On paper, Inkling-Small is an efficiency miracle. A 276B-parameter model with just 12B active parameters scored 40 on Artificial Analysis’ Intelligence Index, one point behind its 975B-parameter sibling Inkling, which activates 41B. It beat that bigger model on SWE-bench Verified and Humanity’s Last Exam. The price is $1.20 per million output tokens, roughly 70 percent cheaper than Inkling. The license is Apache 2.0, the most permissive open-source contract in existence. Mira Murati, former OpenAI CTO, is betting that this combination of performance, price, and openness will make Thinking Machines Lab a household name. But on a blockchain-native reading, the most important number is not 12B or 40. It is 171GB, the size of the quantized weight files. That single figure tells you who can actually run this model, and who cannot. It exposes a governance question that the AI industry has not yet learned to answer. I have spent the last decade watching projects promise decentralization while quietly concentrating control in the hands of a few operators. Inkling-Small is not a cryptocurrency project, but it is running the same playbook. The question is not whether the model is open. The question is who gets to share in the responsibility for how it is used. Thinking Machines Lab is Murati’s post-OpenAI vehicle, and Inkling-Small is its opening salvo. The model is a Mixture-of-Experts multimodal reasoning system: text, image, and audio inputs, but no image or audio generation on the output side. That makes it an understanding model, not a generative one. The company chose a twin-track commercial path: release full weights under Apache 2.0 and sell API access at a price designed to undercut not only its own flagship but much of the frontier API market. The performance data is selective but telling. Compared with Inkling, the small model is no more than one Intelligence Index point behind, while actually reversing the bigger model’s lead on coding and hard reasoning. This is not the typical “small model distilled from a large teacher” pattern. It looks like a deliberately tuned specialist. A 12B-active model does not beat a 41B-active model on SWE-bench by accident. It beats it because the training data allocation, routing, and post-training recipe were pointed at specific tasks. From my experience auditing ICO whitepapers in 2017, I learned that numbers that look too good usually hide a missing denominator. For Inkling-Small, the hidden denominator is data allocation. The fact that a small model outperforms its larger sibling on coding and reasoning strongly suggests it was trained or post-trained with a concentrated diet of code, math, and reasoning tasks, not simply scaled down. That means its advantage is domain-specific. On broad knowledge and factual accuracy, Inkling remains better. So the correct mental model is not “small model equals big model.” It is “specialist model outperforms generalist on the specific tasks where it was pointed.” This is a product decision, not a scientific miracle. The arithmetic of unit intelligence makes the strategy explicit. If Inkling costs around $4.00 per million output tokens—70 percent more than Inkling-Small’s $1.20—then the larger model delivers 41 index points for $4.00, or about $0.0976 per point. The small model delivers 40 points for $1.20, or about $0.03 per point. That is more than a threefold improvement in unit intelligence cost. For a developer building an agent that calls a model millions of times a day, that difference is the difference between a product with healthy margins and a product that dies on unit economics. This is the same accounting that made DeFi composability powerful: when you reduce the cost of a primitive, you expand the set of viable applications. But the open-source framing deserves more skepticism than it is receiving. Apache 2.0 is the gold standard for permissive licensing, and it removes the legal friction that haunts many “open weight” releases. Yet in the crypto world, we have learned the hard way that open code does not equal open governance. A DAO can publish all its smart contract code on Etherscan and still be run by three multisig signers. Similarly, a 171GB weight download is technically accessible to anyone, but practically accessible only to teams with serious infrastructure, GPU memory, and MLOps talent. That is not “everyone can run it.” That is “everyone can sign a click-through license while a few AI labs do the running for them.” People first, protocol second. Always. But a protocol that only a fraction of a percent of the population can operate is not a protocol for the people; it is a protocol for the credentialed. The physical basis of the price is the MoE architecture itself. 975B total parameters suggest a training run on the order of 2×10^25 FLOPs. That implies thousands of H100s for months, and a budget with a staggering number of zeros attached. Yet only 41B parameters are active per token, bringing inference cost down to the range of much smaller dense models. For Inkling-Small, 12B active parameters places it near the middle of the market in serving costs while competing with the top of the market on specialized tasks. That is the engineering trick behind the $1.20 price. Cheap inference is not a marketing gimmick; it is the only strategy that makes sense when your total training bill is too high to recoup through token sales alone. You are not selling tokens. You are buying ecosystem share and usage data. A critical signal is what the release does not say. There is no direct comparison to GPT, Claude, or Gemini. If your model truly competed at the frontier, you would publish that comparison. Avoiding it is itself a data point. Intelligence Index 40 might be a solid score, but on Artificial Analysis’s historical scale, frontier models often sit in the 50-plus range. The “only one point behind” narrative is doing heavy lifting inside a relatively modest band. It is a comparison to a sibling, not to the world. That does not make Inkling-Small unimpressive; it just means the competitive claim is narrower than the marketing suggests. The contrarian take is not that open weights are reckless. It is that open weights, by themselves, are a governance illusion. The AI safety debate has largely been framed as open versus closed. But the real split is between organizations that can operationalize a 171GB model and organizations that cannot. Releasing weights under Apache 2.0 does not decentralize AI; it relocates the centralization point. Instead of concentration in the hands of API providers, we get concentration in the hands of whoever can afford the GPUs and MLOps engineers to fine-tune, serve, and secure the model. That is a smaller club, not a broader one. Even more concerning is the silence around alignment. The materials mention HLE performance, but there is no red-team report, no constitutional alignment methodology, no model card with refusal rates. For a founder whose credibility rests on a safety culture, that silence is a signal. In decentralized finance, we called this “audit theater”: a project pays for a security review but refuses to publish the full report. In AI, withholding safety evals while publishing full weights creates a permanent, irreversible exposure. You cannot recall a 171GB file once it has been mirrored across the world. Trust is earned in bear markets, and the AI industry is entering its own bear market of accountability. This is the moment where trust will either be built or squandered. Empathy is the ultimate security layer. That sentence sounds soft, but it is brutally practical. A model that can transcribe audio, reason about code, and be fine-tuned without oversight can be turned into a tool for voice impersonation, automated social engineering, or stealthy vulnerability discovery. The people who will deploy Inkling-Small are not all benevolent. Open weights are permanent. The only defense is to build governance rails around the ecosystem: provenance registries, fine-tune audit trails, and decentralized mechanisms for reporting misuse. Without those, open weights are not freedom; they are just an unpatched contract. The investment narrative is equally double-edged. Murati’s founder credibility is real, and it can move enterprise customers. But the cost structure is brutal. A 975B-parameter training run, even with efficient MoE, demands capital in the tens of millions of dollars per full cycle. At $1.20 per million output tokens, even hundreds of millions of API calls per day would produce only millions of dollars in monthly revenue, far below the burn rate of a frontier lab. This is not a business model yet. It is a land grab. The strategy is to flood the zone with open weights, capture developer mindshare, and then monetize through enterprise services, fine-tuning, and managed deployment. That is a viable strategy, but it is also a bet that the ecosystem will remember the name when the next round of funding arrives. What would change my confidence? A technical report with training data composition, context window details, throughput benchmarks, and third-party red-team results. Without those, we are left with the same problem that plagued early DeFi audits: beautiful claims and no transparency. I have seen multisig wallets with five signers called “decentralized treasury management.” I have seen governance tokens with no binding power sold as community ownership. Inkling-Small is not a scam, but it is being introduced into a world where openness is a spectrum, not a binary. The question I care about is not whether Inkling-Small is smarter than its bigger sibling. It is whether the community around it can build the governance equivalent of a decentralized treasury: something that gives every user a voice in how the model evolves, how it is fine-tuned, and how misuse is handled. The next frontier is not another 10^25 FLOPs. It is the governance infrastructure for open intelligence. If Thinking Machines Lab treats Apache 2.0 as the finish line, it will be remembered as a solid lab with a good model. If it treats open weights as the starting point for collective oversight, it will define the next era of AI. People first, protocol second. Always.