A New Player Puts Its Cards on the Table
Thinking Machines Lab has released its first AI model, called Inkling, a 975-billion-parameter open source system built to process both video and audio. The release marks the company’s public debut in a market already crowded with well-funded, well-known names – and signals that it intends to compete directly, not occupy some narrow niche.
The scale of the model is not incidental. At 975 billion parameters, Inkling lands in the upper tier of publicly available AI systems, putting Thinking Machines Lab on the same general map as organizations like Anthropic and OpenAI from day one.
Open source, multimodal, and large – the company made three specific bets at once.
What Inkling Actually Does
The defining technical characteristic of Inkling is its multimodal design. Most AI models that have dominated headlines are primarily text-based – they read and generate language, sometimes paired with image understanding. Inkling was trained specifically to understand video and audio, which pushes it into territory that fewer models have staked out with any consistency.
Video and audio comprehension at scale is harder than it sounds. Video requires tracking motion, scene changes, and temporal relationships across frames. Audio adds another layer – tone, speaker identity, ambient sound, language – that pure image models never have to handle. Training a single model to manage all of this, at 975 billion parameters, requires both significant infrastructure and deliberate architectural choices. Thinking Machines Lab has not yet detailed the full training setup or data sources publicly, but the parameter count alone suggests a compute investment that places this well outside hobbyist or small-team territory.
The open source decision also carries weight. Releasing model weights publicly means other developers, researchers, and companies can build on Inkling without licensing fees or API restrictions. That creates adoption pressure on competitors – particularly those who keep their most capable models behind closed APIs – by giving the broader developer community something to work with directly. Whether Thinking Machines Lab plans to monetize through services, cloud infrastructure, or enterprise support on top of the open weights remains an open question.
The Competitive Landscape Inkling Is Walking Into
Anthropic and OpenAI are the obvious reference points, and Thinking Machines Lab’s positioning against them is deliberate enough that calling it implicit would be generous. Both companies have raised billions in funding, deployed widely used products, and spent years building brand recognition in both enterprise and consumer markets. Entering that conversation with a first model – regardless of its size – is a different challenge than releasing a second or third generation system with an existing user base behind it.
Open source has become a legitimate strategic angle against proprietary players. Meta’s Llama series demonstrated that releasing capable open weights could generate massive developer adoption and ecosystem momentum faster than any marketing campaign. Mistral built an entire company identity around open, efficient models. Thinking Machines Lab appears to be reading from a similar playbook, though with a specific focus on video and audio comprehension rather than pure language efficiency. That specialization could help it carve out concrete use cases – video analysis, transcription-adjacent tasks, multimedia content processing – where existing open models offer less.
The risk is that 975 billion parameters, while large, does not guarantee quality. Parameter count has repeatedly proven to be an imperfect proxy for actual capability. What matters in practice is how well Inkling performs on real tasks – whether its audio understanding is reliable enough for production use, whether its video comprehension can handle edge cases, and whether developers find it worth the infrastructure cost to run a model of that size. Benchmarks will follow, and the results will matter more for Thinking Machines Lab’s competitive standing than the announcement itself.
Why This Release Matters Beyond the Model Itself
Thinking Machines Lab is now on record. The company has shipped something with its name on it, and that changes the nature of the organization from a known quantity inside AI circles to a public actor with a product that can be downloaded, tested, critiqued, and compared. That shift – from potential to actual – is where most AI startups either build credibility or expose gaps.
The multimodal angle, specifically the combination of video and audio in a single open model at this scale, is not something every lab has prioritized. Text-and-image has become the standard pairing for multimodal releases. Adding video and audio comprehension as the primary design goal, rather than as a secondary feature bolted onto a language model, suggests Thinking Machines Lab made a deliberate call about where it thought the opportunity was underserved. Whether that bet reflects a genuine gap in the market or a misread of where demand is actually concentrated will become clear as developers begin working with Inkling directly.
The company’s ability to follow Inkling with iterative improvements – and to do so at a pace that keeps it relevant as Anthropic, OpenAI, and others continue shipping – is the real test. First models establish presence. Second and third models establish trajectory.
Inkling is currently available as an open source release, with 975 billion parameters trained across video and audio modalities. The developer community’s reception over the next several weeks will determine whether Thinking Machines Lab’s entry into this market is remembered as a strong opening or simply a large number attached to an unfamiliar name.
