The former CTO of OpenAI just bet $12 billion that enterprises don't want their AI models polite—they want them honest, cheap, and private.
The Summary
- Thinking Machines released Inkling, a 975-billion-parameter open-source multimodal model under Apache 2.0 license, designed to run on-premises with "resistance to censorship"
- The model scores 77.6% on SWE-bench Verified (beating Nvidia's Nemotron 3) and handles text, images, and audio natively
- Mira Murati's startup is valued at $12 billion after 18 months building AI infrastructure out of public view
- Inkling features "controllable thinking effort" to let enterprises dial cost against performance—the first major model explicitly built for factual output regardless of controversy
The Signal
Thinking Machines isn't playing the chatbot game. While Anthropic and OpenAI compete on who can make Claude or GPT sound more helpful and harmless, Mira Murati's new company just shipped a model designed to give enterprises what safety-obsessed API providers won't. The company explicitly built Inkling "to answer directly on topics that may be subject to censorship", positioning it as the option for companies that need factual outputs on sensitive topics without running into content policies written by liability-averse legal teams.
This matters because the agent economy runs on reliability, not politeness. If your AI agent handles procurement, legal research, or risk assessment, you need answers about sanctioned entities, controversial suppliers, or politically sensitive markets. You need those answers fast, cheap, and on your own infrastructure where compliance teams can audit every query. Inkling's Apache 2.0 license and availability on Hugging Face means enterprises can customize it, run it in private clouds, and never send a single query to someone else's API.
"Enterprises looking to move more of their agentic AI workloads to open weights models they can customize, control and run on-premises have a strong new contender."
The technical specs back up the positioning. As a Mixture-of-Experts architecture, Inkling activates only the parameters it needs for each task, keeping inference costs down. The "controllable thinking effort" mechanism lets developers trade accuracy for speed, something closed models don't expose. Compare the numbers:
- 77.6% on SWE-bench Verified vs. Nvidia Nemotron 3's 71.9% on coding tasks
- 91.4% on VoiceBench for voice understanding (trailing Gemini 3.1 Pro's 94.4%)
- Native multimodal processing of text, images, and audio without stitching together separate models
The timing is strategic. While U.S. regulators debate export controls and China floods the market with free Qwen and DeepSeek models, Thinking Machines positions itself as the American open-weights champion. The $12 billion valuation before shipping a single model suggests investors believe enterprises will pay for sovereignty—control over where models run, what they say, and who sees the queries. That's the real product here. Inkling isn't trying to be the best model. It's trying to be the model enterprises trust when they can't afford to trust anyone else's API.
The Implication
Watch how fast this becomes infrastructure. If enterprises adopt Inkling for sensitive workflows, Thinking Machines builds a moat in the most valuable part of the market: places where control matters more than leaderboard scores. The agent economy doesn't run on the best model. It runs on the model you can customize, audit, and trust to run the same way at 3 a.m. six months from now. For builders, the takeaway is simple: Apache 2.0 means you can actually build on this. For everyone else, notice how Murati is attacking OpenAI from the opposite direction—not more closed, more open. Not safer guardrails, fewer guardrails. That's the bet. We'll know if it paid off when we see which Fortune 500 companies stop sending queries to Sam Altman.
Sources
VentureBeat | Fortune Tech | Wired AI | TechCrunch AI | Bloomberg Tech | Hugging Face Blog