Thinking Machines Lab Launches Inkling, a 975B-Parameter Open-Weight Model Built by OpenAI Defectors

The Release

Thinking Machines Lab, a startup founded by former OpenAI executives including Mira Murati (former CTO) and John Schulman (ChatGPT co-developer), has launched Inkling, a 975-billion-parameter open-weight model designed for multimodal reasoning. According to Wired AI, the model processes audio, video, and text inputs and performs competitively on advanced reasoning and coding tasks, though the company acknowledges Inkling does not rank at the top of popular benchmarks. The release marks the startup’s entry into the crowded open-weights space, positioning it against Chinese models and closed-source systems from OpenAI and others.

Emergent Reasoning Optimization

A noteworthy finding emerged during Inkling’s training: the model’s chain-of-thought explanations became progressively more concise without sacrificing comprehension or final outputs. According to the company’s technical documentation cited by Wired AI, this optimization dropped grammatical overhead while maintaining reasoning integrity. This observation suggests that models can self-optimize their computational pathways when trained with appropriate signals—a phenomenon that could have implications for inference efficiency in production systems.

Strategic Positioning in the Open-Weights Landscape

The release reflects a deliberate business strategy to establish Thinking Machines Lab as a credible player in the decentralized AI movement. Wired AI reports that the company emphasizes AI should not be controlled by a small number of corporations, advocating instead for distributed development powered by open-weights models. Open-weights systems offer cost and customization advantages over closed APIs, though they require substantial compute infrastructure—Inkling requires clusters of specialized chips to operate efficiently. The company previously released Tinker, a fine-tuning tool, and voice-interaction capabilities, signaling a product ecosystem approach beyond the base model.

Defector-Led Competition

Thinking Machines Lab exemplifies a broader pattern where departing executives from market leaders establish competitive ventures. According to Wired AI, the startup was valued at $12 billion in its seed round—the largest seed valuation on record—reflecting investor appetite for teams with proven AI expertise. Anthropic, another OpenAI-defector company, is reportedly preparing an IPO that could value it above $1 trillion, with its Claude model gaining adoption for coding and business applications. This wave of departure-driven startups is reshaping the competitive structure of the AI industry, fragmenting what was once dominated by OpenAI’s incumbency.

Why This Matters

For teams evaluating model infrastructure, Inkling’s release expands the viable open-weights options beyond Chinese models, offering US-trained alternatives that do not carry the same geopolitical sourcing concerns. For organizations building proprietary AI systems, the availability of a 975B-parameter open-weights baseline reduces the technical moat of closed models—custom fine-tuning and domain adaptation are now accessible without API dependencies. However, the model’s size means it is inaccessible to resource-constrained teams; whether Thinking Machines Lab will release smaller quantized variants will determine how broadly it disrupts the market. The company’s demonstrated ability to attract top talent and raise record funding suggests it can sustain long-term competition, but execution on multiple product lines—not just base models—will determine whether it becomes a durable alternative to OpenAI or remains a well-funded but niche player.