On March 15, 2025, a new model appeared on OpenRouter. Its name: Inkling. Its pedigree: Mira Murati, former OpenAI CTO. Its claim: the best Western open-source model. Within hours, the hype machine spun up. The code never lies, only the auditors do. But the code wasn't there. This is not a review. This is an autopsy.
Context
Thinking Machines Lab emerged from stealth two years after Murati's departure from OpenAI. The team promised a paradigm shift: an AI model that mastered the Model Context Protocol (MCP), the so-called standard for agent tool calling. The announcement landed on a blockchain and Web3 news site—a curious venue for an AI project. The lack of technical detail was striking: no paper, no GitHub repository, no benchmark scores. Just a single metric: 'impressive MCP scores.' This is the same pattern I saw in 2017 ICOs: a whitepaper full of vision, empty of code. Complexity is just laziness wearing a tech suit.
Core
Forensics reveal the truth markets try to bury. Let's dissect what we know. The only measurable claim is 'MCP score.' MCP is not a standard benchmark like MMLU or HumanEval. It's a protocol for model-context interaction, primarily relevant to agentic workflows. Emphasizing MCP is like a DeFi project bragging about its gas optimization while ignoring security audits. It's a narrow win. The model's architecture? Unknown. Parameter count? Unknown. Training data? Unknown. Even the open-source license remains unconfirmed. In my 2017 code audits, I learned to flag projects that hide their math. Inkling is hiding its math.
Let's stress-test the 'best Western open-source' claim. Best by which metric? Compare with Llama 3.1 405B, Mistral Large, DeepSeek-V3. These models have public benchmarks, known architectures, and verifiable weights. Inkling has none of that. The phrase 'Western' is a tell—it deliberately excludes superior Eastern models like Qwen or DeepSeek. This is marketing, not engineering. The code never lies, but the marketing copy does.
The MCP emphasis suggests a post-training optimization, not a base-model breakthrough. Most likely, Inkling is a fine-tuned derivative of an existing open-source foundation—probably Llama or Mistral. The team spent two years on alignment and agent-specific tuning. That's plausible. But it's not innovation; it's specialization. The hidden question: how much of the training data came from OpenAI's internal systems? Murati's exit from OpenAI may have included non-disclosure agreements, but the talent she took likely carried unpatented knowledge. This is a regulatory gray area, not a technical moat.
From a safety perspective, an agent-focused model is exponentially more dangerous than a chatbot. One prompt injection can lead to a cascade of malicious tool calls. The announcement was silent on alignment mechanisms. Red teaming? Unknown. Constitutional AI? Unknown. In my 2025 regulatory analysis, I found that 40% of DeFi protocols failed basic KYC checks. Inkling's safety posture is even less transparent. Tracing the silent bleed from 2017's broken logic: we are repeating the same mistakes.
Contrarian Angle
What did the bulls get right? First, the team's pedigree is genuine. Murati's reputation for safety and alignment at OpenAI is well-documented. If Inkling truly advances MCP, it could establish a de facto standard for agent interoperability—much like ERC-20 did for tokens. Second, the two-year silence could indicate a deliberate, high-quality build, not a lack of substance. Third, OpenRouter as a launchpad is intelligent: it grants access to a developer audience without the overhead of a proprietary platform.
But these are hypotheses, not proofs. The bulls confuse reputation with evidence. Murati's past does not guarantee her present. The MCP standard might never achieve critical mass, especially against entrenched frameworks like LangChain or Anthropic's Tool Use. The silence might also hide failure: teams that build in secret often emerge with nothing. Forensics reveal the truth, but only when the evidence is available. Currently, the evidence is missing.
Takeaway
The market will price in hype. Forensics price in truth. Until the weights are public, the benchmarks are verifiable, and the safety audits are shared, Inkling is just a promise. Promises don't compile. The community should demand a public repository, a technical paper, and independent third-party testing. Without these, this announcement is noise—designed to capture attention before a funding round. Tracing the silent bleed from 2017's broken logic: we have seen this play before. The code never lies, only the auditors do. And in this case, there are no auditors.