Thinking Machines, the startup led by former OpenAI CTO Mira Murati, released Inkling Small, an open-source model that matches much of its flagship’s performance at roughly a quarter of the size.
The 276-billion-parameter multimodal reasoning model scores 40 on the Artificial Analysis Intelligence Index, one point behind the 975-billion-parameter Inkling, and actually beats the larger model on several coding and reasoning benchmarks, including SWE-bench Verified and Terminal Bench 2.1.
The gap comes from architecture. Inkling Small is a sparse mixture-of-experts model with 256 experts, routing each token through six of them plus two shared experts, for 12 billion active parameters per token versus 41 billion in the flagship. It handles text, image and audio inputs, supports a million-token context window, and ships under an Apache 2.0 license with full weights on Hugging Face.
For enterprises the pitch is economics: near-flagship capability with a fraction of the compute and inference cost. The company is offering a limited-time 50% discount that puts standard 64K-context pricing at $0.58 per million prefill tokens and $1.44 per million output tokens. A 256K-context variant costs more.
The release follows the original Inkling debut by just two weeks and gives Thinking Machines a fast-moving small-model line as the open-weight price war intensifies. The company also added fine-tuning support through its Tinker training API.
