What's this about?
Thinking Machines — the AI lab founded by former OpenAI CTO Mira Murati — released Inkling, a model they trained from scratch and published with full weights available for anyone to download. It's huge (975 billion parameters, though only 41B "active" per token), handles text, images, and audio natively, supports a 1M-token context window, and was trained on 45 trillion tokens. A lighter sibling, Inkling-Small, ships as a preview. Notably, the lab is honest that it's *not* the strongest model available — it's meant to be a good base for customization via their Tinker fine-tuning platform.
What exactly are "open weights"?
A model's weights are the billions of learned numbers that make it work — the actual "brain," not just the app around it. Most top models (GPT, Claude, Gemini) keep weights secret; you can only rent access through an API. Open weights means the file is published so anyone can download, run, inspect, and fine-tune it on their own hardware — like getting the master recording instead of a streaming subscription. Until now, the best open-weights models have almost all come from Chinese labs (DeepSeek, GLM, Kimi).
What's the community saying?
- "America finally has its own DeepSeek" — many root for Thinking Machines as the US answer to Chinese open-weights dominance; several call it the best American open-weights model to date.
- But is it actually good? Skeptics note it's ~30% bigger yet weaker than China's GLM 5.2, and debuts mid-table on the Artificial Analysis index after a $2B raise — "what a joke," per one commenter; defenders reply it's a remarkably strong *first* model and clearly a foundation to build on.
- The audio angle stands out — it's the largest open-weights model with native audio; llama.cpp and Unsloth support appeared within hours.
- The business-model question — how do you make money giving weights away? Debate over whether Tinker fine-tuning is the real product.