Thinking Machines Lab, the closely watched startup founded last year by former OpenAI CTO Mira Murati, released its first in-house AI model on July 15. Called Inkling, the mixture-of-experts (MoE) system carries 975 billion total parameters and activates just 41 billion for any given token, a design that keeps inference cost far below dense frontier competitors of similar quality.
Open-weight bet against one-size-fits-all AI
Unlike closed flagships from OpenAI, Anthropic and Google, Inkling ships with open weights and adapter-friendly tooling. Murati's team is explicitly positioning Inkling as a base for enterprise fine-tuning across domains, from code assistants to specialised scientific workflows, arguing that customer-owned checkpoints beat monolithic APIs for hard, regulated deployments.
$2B war chest, 41B active parameters
Reports place Thinking Machines' most recent private round at roughly $2 billion at a $12 billion valuation, backed by Andreessen Horowitz and Nvidia. The lab is spending the capital on both training runs and researcher recruitment, with Inkling positioned as the first stop in a family of open-weight models. Head-to-head benchmarks against Meta Llama 4 and Alibaba Qwen 3 are expected in the coming weeks.
Enterprise fine-tuning gets a new baseline
Inkling lands into an increasingly crowded open model market. It follows Moonshot's 2.8T Kimi K3 release and Google's Gemini 3.5 Pro debut, and comes as customers weigh the case for multi-model compute deals. Analysts say the open MoE approach will pressure per-token API pricing.
Reporting based on coverage from TechCrunch, Reuters and Thinking Machines Lab.