AI Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

• Inkling is Thinking Machines Lab’s first open-weights, multimodal model — a 975-billion-parameter, Mixture-of-Experts architecture trained to understand video and audio, with controllable reasoning effort [Wired][ThinkingMachines].
• The model is released as open weights so enterprises can customize and fine-tune it (via Thinking Machines’ Tinker platform) rather than rent closed, one-size-fits-all systems from frontier labs [ThinkingMachines][Wired].
• The launch positions Thinking Machines to compete with major AI labs by offering a customizable alternative and broader access to a high-capability multimodal model [Wired].

Follow-up Questions:

1. How does Inkling’s Mixture-of-Experts (MoE) architecture differ from dense models in practice?

2. What are the licensing and commercial-use terms for Inkling’s open weights?

3. What datasets and safety mitigations did Thinking Machines use when training Inkling?

4. How does Inkling’s performance compare to current OpenAI/Anthropic models on standard benchmarks?

5. How can an enterprise get access to Tinker for fine-tuning Inkling?

Sources: [ThinkingMachines], [Wired]