AI startup Thinking Machines has launched Inkling, its first publicly available model, positioning it not as a frontier competitor but as a flexible, open-weight base for enterprise and developer fine-tuning. The model is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters, supporting a context window of up to one million tokens.
Pretrained on 45 trillion tokens spanning text, images, audio, and video, Inkling offers native multimodal reasoning across all three input types — a capability that distinguishes it from most open-weight alternatives, which typically lack native audio support. Full weights are available on Hugging Face, and fine-tuning is accessible via the company’s Tinker platform.
The company is transparent about Inkling’s positioning: it does not claim state-of-the-art status across the board. Benchmark results show competitive but not leading performance compared to closed-weight models such as Claude Fable 5 and GPT-5.6 Sol on reasoning and agentic tasks.
Instead, the release emphasizes breadth — strong performance across coding, instruction following, factuality, vision, and audio — alongside a key differentiator: controllable thinking effort. Developers can tune how many tokens the model uses to solve a problem, enabling significant cost and latency savings. In testing, Inkling matched Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one-third the token cost.
Today, we are introducing Inkling.
— Thinking Machines (@thinkymachines) July 15, 2026
Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.https://t.co/Ghebq5mG30
Available today for fine-tuning on Tinker. Play with it in the Inkling Playground.
Beyond raw capability, Thinking Machines invested considerably in the model’s epistemic behavior and safety profile. Inkling was trained using reinforcement learning against proper scoring rules on a large corpus of resolved real-world forecasting questions, producing a model calibrated to express appropriate uncertainty rather than confidently hallucinating.
On ForecastBench, it performs on par with leading closed models including Gemini 3.1 Pro and Grok 4.3. The training pipeline also incorporated dual automated graders — a rubric grader and a claims grader with agentic web search — to simultaneously improve helpfulness and reduce factual errors.
On safety, Inkling leads open-weight models on FORTRESS, a benchmark evaluating refusal of harmful requests while avoiding over-refusal of benign analogs, scoring 78% on adversarial prompts against 77.6% for Nemotron 3 Ultra and 65.6% for Kimi K2.6.
Alongside Inkling, Thinking Machines previewed Inkling-Small, a lighter 276B-parameter model with 12B active parameters that matches or exceeds the larger model on several benchmarks, offering a lower-cost option for synthesis and grading workloads. Both models are currently available through Tinker, with deployment partnerships spanning TogetherAI, Fireworks, Databricks, Hugging Face, and others.
The post Thinking Machines Debuts Inkling: A Foundation Model Optimized For Customization appeared first on Metaverse Post.
