Inkling, the 975B-Parameter Question OpenAI Wouldn’t Answer
Mira Murati left OpenAI in late 2024. Eighteen months later, her new company — Thinking Machines Lab — just dropped a 975-billion-parameter open-weights model called Inkling. Under Apache 2.0. Weights available. Go download it.
That’s something OpenAI has never done and, as far as anyone can tell, never will.
Let me be direct about what this means. China has been eating America’s lunch on open frontier models for over a year now. DeepSeek V4, GLM 5.2, Kimi K2.6 — these are serious models, released with weights you can actually touch. American labs have been keeping their best stuff behind API walls, selling tokens instead of sharing knowledge. Inkling is the first American open-weights model that competes in that weight class.
The numbers. 975B total parameters, 41B active per token (mixture-of-experts, 256 routed experts + 2 shared ones). 45 trillion tokens of training data — text, images, audio, video. A million-token context window. Trained on Nvidia GB300 NVL72 systems from scratch (they credit DeepSeek-V3 for architectural inspiration, but it’s their own training run). It matches Nvidia’s Nemotron 3 Ultra (the previous American open-weights champ at 550B params) on Terminal Bench 2.1 using a third of the thinking tokens.
It’s not the best model on the market. Their own benchmarks show it trailing Claude and GPT on several tasks. But that’s not the point.
The point is ownership. When you run a model through an API, you’re renting cognition. When someone sends you a fine-tuned model, you’re inheriting someone else’s judgment. Open weights mean you can inspect what the model actually learned. You can fine-tune it yourself. You can run it on your own hardware. You can verify it’s not doing anything weird with your data.
Inkling even wrote its own fine-tuning scripts — the company showed it using their Tinker platform to self-retrain. That’s a flex, but it’s also a genuinely useful capability for anyone who wants to customize the thing without being a machine learning PhD.
The hardware problem is real. You need about eight Nvidia B300s or sixteen H200s to run the full 16-bit model. That’s a datacenter, not a homelab. The NVFP4 quantized version cuts that in half, but you’re still looking at serious infrastructure. You can also just use the API through Tinker or partners like TogetherAI, Fireworks, and Databricks.
The counterargument I keep hearing: “What’s the point of open weights if nobody can run them?” Fair question. But it misses the real value. Open weights let you audit. They let you fine-tune. They let you build on top. The API is the on-ramp; the weights are the insurance policy. You don’t need to run a 975B model at home to benefit from it being open — you need to know that what you’re paying for isn’t a black box with a company’s thumb on the scale.
The bigger story here is about talent leaving OpenAI and doing what OpenAI won’t. This isn’t just Murati. Ilya Sutskever’s Safe Superintelligence Inc. is cooking something. John Schulman’s Anthropic is the closest thing to a lab that takes safety seriously. The diaspora is real, and it’s producing more interesting work than the mothership.
The objection I wrestle with: “Murati is just proving she can build what OpenAI built, only worse.” I don’t buy it. She’s proving that frontier AI doesn’t have to be a walled garden. The Apache 2.0 license is the signal. If you believe — as I do — that AI safety and AI capability both benefit from transparency, then Inkling is a step forward even if it doesn’t top every leaderboard.
OpenAI launched with a nonprofit mission to build AGI that benefits everyone. Somewhere along the way, that became “build AGI that benefits OpenAI’s valuation.” Murati’s Inkling isn’t the completion of that original mission. But it’s the first real evidence in a long time that someone still remembers what it looked like.
Inkling won’t change the world by being the smartest model in the room. It’ll change it by being the model you can actually own.
Sources: The Register - Former OpenAI CTO releases open frontier model, Thinking Machines Lab - Inkling announcement, Hacker News discussion (733 points)