Inkling: An Open-Weight Rebuttal to One-Size-Fits-All AI
Inkling is a 975-billion-parameter open-weight Mixture-of-Experts AI model designed for advanced text, audio, and video understanding, released by Thinking Machines Lab to give developers and organizations direct control over customization and deployment while challenging closed, proprietary general-purpose systems. That design choice is not a side detail; it is the point. In a market dominated by monolithic models sold as finished products, Inkling is framed as a starting layer that enterprises reshape with their own tools, data, and constraints. Instead of paying to query a sealed model from OpenAI or Anthropic, teams can download Inkling’s weights, adjust its behavior, and even integrate it as the reasoning core behind their own products. The message is blunt: AI should be something you build with, not only something you rent access to.
Inside a 975B-Parameter Mixture-of-Experts Built for Media
Inkling is not just large; it is architected for specialization. It is a Mixture-of-Experts transformer with 975 billion total parameters and about 41 billion active during inference, letting different expert subnetworks handle different tasks without lighting up the entire model each time. It was trained from scratch on 45 trillion tokens across text, images, audio, and video, giving it native multi-modal AI capabilities rather than bolted-on adapters. Practically, that means a single open-weight AI model can reason over documents, analyze frames from video, and understand long-form audio, then respond in text. It also supports context windows up to one million tokens, pushing it firmly into territory where continuous streams—earnings calls, medical recordings, compliance footage—can be processed in a single session. You either see this as overkill, or as the baseline for serious enterprise work.
Customization as Strategy, Not Feature
Thinking Machines is betting that customizable AI deployment will beat polished, closed models in the places that matter: cost, control, and fit to domain. Inkling is available as an open-weight foundation model, meaning developers and companies get full access to download, fine-tune, and deploy it. Their Tinker platform wraps this with 64K and 256K context options and discounted fine-tuning, explicitly positioning Inkling as infrastructure for custom systems rather than an end-user assistant. One quotable example backs up the thesis: in a project with Bridgewater Associates, a fine-tuned open-source model reportedly outperformed proprietary systems on financial reasoning tests at a fraction of the cost. This is the quiet revolt against per-token pricing and vendor lock-in; if open-weight AI models can match or exceed closed models on specific workloads, executives will start asking why they are paying to use what they could own and adapt.
Real Capabilities Today, Not Just Ideology
Inkling would be easy to dismiss as an ideological project if it were weak. It is not. The model handles reasoning, coding, tool use, instruction following, visual analysis, speech transcription, and long-form audio understanding out of the box. It can run inside coding-agent harnesses, cope with changing tool schemas, and iteratively produce structured artifacts and applications. On benchmarks, it scored 77.6% on SWE-bench Verified, 97.1% on AIME 2026, 87.2% on GPQA Diamond, and 73.5% on MMMU Pro. According to the company, Inkling can match Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly one-third as many generated tokens. Users can even control how hard the model “thinks” by dialing effort between 0.2 and 0.99 to trade off speed, quality, and token usage. This is a multi-modal AI system that respects practical constraints instead of chasing leaderboard glory alone.
A Credible Challenger with a Long Game
Inkling also matters because of who is behind it. Thinking Machines Lab was founded by former OpenAI CTO (and brief CEO) Mira Murati, alongside other senior OpenAI veterans John Schulman and Lilian Weng. The company launched in February 2025 with the largest seed round in history, valued at USD 12 billion (approx. RM55.2 billion) from the outset, and now employs around 200 people. It has already shipped a fine-tuning tool, natural voice interaction systems, and machine-learning research, and used Inkling to refine itself during training. The lab is previewing Inkling-Small, a 276-billion-parameter Mixture-of-Experts model with 12 billion active parameters that approaches or surpasses Inkling on several tasks while targeting lower cost and lower latency, with full weights promised after testing. The release marks the start of a planned family of customizable models across sizes. In a landscape where closed giants like OpenAI and Anthropic now dominate attention, “exiles” shipping serious open-weight systems is more than symbolism—it is competitive pressure.
