Together AI brings Thinking Machines Lab’s new model Inkling on day 0
Together AI ● Covered by 2 sources
Thinking Machines Lab just dropped Inkling, a huge new multimodal AI model, and Together AI already has it running.
Based on reporting by Together AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Thinking Machines Lab, the startup that's been quietly building toward something big, finally shipped its first major model today: Inkling. And Together AI didn't wait around to offer it — the model went live on their serverless platform the same day, which in this industry counts as fast even by Silicon Valley standards.
Inkling is a mixture-of-experts model with 975 billion total parameters, though only 40 billion activate per token, paired with a 1 million token context window. It takes text, images, and audio all through one decoder stack and spits out text. No separate pipelines, no bolted-on vision encoder treated as an afterthought — the embeddings for images and audio get squeezed into the same width as text tokens and fed straight into the model alongside everything else.
What's more interesting than the size is the architecture underneath. Instead of the usual RoPE position embeddings most transformer models lean on, Inkling uses a learned, query-conditioned relative bias baked directly into attention. It also mixes sliding-window attention with occasional full-context layers, five to one, and throws in something called sconv — short causal convolutions layered before and after attention to catch nearby-token relationships cheaply. The mixture-of-experts routing has a shared expert sink too, which normalizes shared and routed experts together rather than treating the shared path as a fixed baseline. None of this is standard-issue transformer plumbing.
Thinking Machines also built in controllable inference effort, so developers can dial reasoning depth up or down depending on whether they need speed or depth for a given task. Together AI says it built a custom FlashAttention-4 kernel specifically to handle Inkling's relative-attention mechanism at production scale, which suggests this wasn't a trivial integration job. Early benchmark numbers reportedly show strength across scientific reasoning, competition math, coding, and even calibrated forecasting — a less common training target that matters for anyone building systems that need to express uncertainty rather than just guess confidently.
My take — AI-written commentary, not fact-checked reporting
A 975-billion-parameter MoE with novel attention math shipping day-zero on a third-party inference platform is the kind of launch that used to take months of waiting lists — so credit to Together AI for the plumbing, but I'd hold my applause on Inkling itself until independent benchmarks show up, because self-reported 'strong preliminary evaluations' from a lab's own blog post is marketing, not proof.
Read more about this at: Together AI