Inkling: Thinking Machines Lab's Open-Weights Multimodal Base Model
— by Vivax
Thinking Machines Lab has released Inkling, an open-weights Mixture-of-Experts transformer with 975 billion total and 41 billion active parameters, a…
On July 15, Thinking Machines Lab introduced Inkling — and with it, one of the most substantial open-weights releases the field has seen. Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, of which 41 billion are active for any given token, pretrained on 45 trillion tokens spanning text, images, audio and video and equipped with a one-million-token context window. Alongside the flagship, the lab is previewing Inkling-Small, a lighter sibling with 12 billion active parameters for teams that need the same recipe at a fraction of the serving cost. Both ship as open weights: you can download the model, inspect it, and run it on your own hardware.
What sets Inkling apart from most open releases is that multimodality is not bolted on. The model reasons natively over text, images and audio inside one architecture, rather than routing each modality through a separate encoder and stitching the results together. It also exposes controllable thinking effort: the same weights can answer quickly with shallow reasoning or spend substantially more computation deliberating on a hard problem, and the caller decides where on that dial each request sits. For anyone building products on top, that single knob — latency and cost on one end, depth on the other — is one of the most practical features in the release.
The positioning is refreshingly honest. Thinking Machines Lab does not claim Inkling is the strongest model in the world; it is presented as a broad, balanced open-weights base built for customization — solid across knowledge, reasoning, code and multimodal understanding, without over-optimizing for any single leaderboard. Where it does stand out is agentic work: Inkling ranks among the strongest open-weights models on agentic coding benchmarks such as Design Arena, which measures how well a model plans, edits and iterates across a real codebase rather than how well it completes a snippet. A balanced base with strong agentic behavior is exactly the profile you want if your plan is to adapt the model, not just to chat with it.
Customization is the point, and the launch makes that concrete. Inkling is fine-tunable on Tinker, the lab's training API, which gains an Inkling Playground for experimenting with the model before committing to a run. The launch demo is delightfully self-referential: Inkling uses Tinker to fine-tune itself into a lipogram model that never uses the letter 'e' — writing fluent prose under a constraint that would trip up most humans. The stunt lands a serious argument: if a model can rewire something as deep as its own lexical distribution through fine-tuning, then adapting it to a specialized vocabulary, a regulated domain or an institutional style is well within reach.
That is why open-weights bases matter more than leaderboard winners. A frontier model behind an API is a service you rent on someone else's terms — the weights can change under you, the data must travel to the provider, and deep adaptation is off the menu. An open base like Inkling inverts all three: the weights are yours to host, audit and specialize, the data never has to leave your infrastructure, and fine-tuning can reshape the model's behavior at a depth no system prompt can reach. For domains with strict data-sovereignty and privacy constraints — medicine above all — that difference is not a preference; it is the precondition for building at all.
This is precisely the thesis Vivax is built on. We are an AI lab building foundational clinical world models — the backbone — with mini apps on top, and a national medical LLM is one of them: not a general chatbot with a medical veneer, but a strong open base adapted deeply to the clinical domain, to Turkish healthcare data, and to infrastructure that hospitals actually control. Releases like Inkling expand exactly the raw material that strategy depends on. The stronger and more customizable the open bases become, the faster grounded, domain-specialized medical AI arrives. The weights are on Hugging Face, the announcement is at thinkingmachines.ai, and the fine-tuning on-ramp is Tinker — the building blocks are on the table.