Inkling: Our Open-Weights Model

1089 points · 269 comments on HN · read original →

Points and comments are a snapshot, not live.

Thinking Machines released Inkling, an open-weights 975B MoE model trained from scratch.

Inkling is a mixture-of-experts transformer with 975B total parameters (41B active) and a 1M-token context window, pretrained on 45T tokens of text, images, audio, and video. It supports native multimodality, controllable thinking effort, and is designed as a base for fine-tuning on the company's Tinker platform. Benchmarks show competitive agentic coding, reasoning, safety, and calibration scores. A smaller 12B-parameter variant, Inkling-Small, is also available.

What commenters are saying

Many commenters root for a US open-weights model to compete with Chinese offerings like DeepSeek and GLM, suggesting Thinking Machines could fill that role. Skeptics note Inkling trails GLM 5.2 on agentic benchmarks and question the business model for open-weights companies, pointing to Tinker's fine-tuning hosting as a potential revenue stream. Several mention other US contenders like Arcee and Reflection but note their models are not yet competitive at this scale.