Thinking Machines finally ships, and it's 975B of open weights
Mira Murati's lab has been the loudest silence in AI for a year. Raised billions, hired half of OpenAI's alignment and research bench, published a few blog posts, shipped nothing you could download. That ended today. Inkling is out, and it's not a toy.
It's a Mixture-of-Experts transformer, 975 billion total parameters, 41 billion active, with a 1 million token context window. Pretrained on 45 trillion tokens across text, images, audio and video, then hammered with more than 30 million reinforcement-learning rollouts in post-training. The architecture skips the usual RoPE for relative positional embeddings and interleaves sliding-window with global attention at a 5-to-1 ratio, which is a bet on long context actually working rather than just fitting in the window.
The numbers are frontier-adjacent. AIME 2026 at 97.1, GPQA Diamond at 87.2, SWEBench Verified at 77.6. That SWEBench figure is the one agent builders should stare at, because it means this thing can close real GitHub issues, not just answer trivia. And the whole thing is open weights on Hugging Face, fine-tunable on their Tinker platform, and served through Together and Fireworks if you don't want to host a 975B model yourself.
Here's why it matters beyond one more model on the leaderboard. A top-tier, ex-OpenAI lab just put a near-frontier multimodal model into the commons, with a knob for how hard it thinks and calibrated uncertainty baked in. The open-weight frontier is no longer just the Chinese labs and Meta. Murati chose to compete by giving the weights away, and that reshapes who gets to build serious agents on top of a model they fully control.
Details and weights at https://thinkingmachines.ai/news/introducing-inkling/
← Back to all articles
It's a Mixture-of-Experts transformer, 975 billion total parameters, 41 billion active, with a 1 million token context window. Pretrained on 45 trillion tokens across text, images, audio and video, then hammered with more than 30 million reinforcement-learning rollouts in post-training. The architecture skips the usual RoPE for relative positional embeddings and interleaves sliding-window with global attention at a 5-to-1 ratio, which is a bet on long context actually working rather than just fitting in the window.
The numbers are frontier-adjacent. AIME 2026 at 97.1, GPQA Diamond at 87.2, SWEBench Verified at 77.6. That SWEBench figure is the one agent builders should stare at, because it means this thing can close real GitHub issues, not just answer trivia. And the whole thing is open weights on Hugging Face, fine-tunable on their Tinker platform, and served through Together and Fireworks if you don't want to host a 975B model yourself.
Here's why it matters beyond one more model on the leaderboard. A top-tier, ex-OpenAI lab just put a near-frontier multimodal model into the commons, with a knob for how hard it thinks and calibrated uncertainty baked in. The open-weight frontier is no longer just the Chinese labs and Meta. Murati chose to compete by giving the weights away, and that reshapes who gets to build serious agents on top of a model they fully control.
Details and weights at https://thinkingmachines.ai/news/introducing-inkling/
Comments