Together AI Adds Inkling Model for Smarter Reasoning
A multimodal setup that lets you tweak how hard the model thinks
@togethercomputeTL;DR:
- Inkling handles text, images and audio in one model
- You can dial inference effort up or down depending on the task
- It runs on FlashAttention-4 for better speed and lower costs
Inkling from Together AI
Together AI is now offering Inkling, a model from Thinking Machines Lab. It's a multimodal mixture-of-experts setup designed for reasoning tasks without wasting tokens.
What it does
Inkling handles text, images, and audio right out of the box. You can adjust how much effort it puts into inference, trading off speed for deeper answers. They optimized it with FlashAttention-4 to run faster.
Why it matters
Being able to tune the effort level helps teams match the model to specific jobs - quick answers or careful ones. Since it works with multiple data types, it could handle documents, image questions, sound inputs, or mixed agent workflows. The MoE design only runs the parts it needs, which keeps costs down when you're doing lots of queries. This also points to where things are heading: platforms that give you good tools and optimized kernels alongside the models.
Quick take
Medium impact for now. It's a model drop, useful for devs, and shows some interesting tech choices.