Inkling 2 Small Leads Open Models on DesignArena

The compelling story is not just benchmark performance—it is a model optimized for long-horizon agent loops and customization. The 15× cost advantage could make frontier-style coding and design agents dramatically more practical, though “local soon” still means waiting for the weights and substantial hardware.

• –Its 1,340 DesignArena score ranks third overall and first among open-weight models in the announcement’s snapshot.
• –The model was trained on hundreds of thousands of long-horizon coding tasks with visual feedback, targeting real agent workflows rather than chat alone.
• –Tinker access is available now, but the model remains closed-weight initially; Apache 2.0 local deployment is planned after external red-teaming.
• –The cost claim is most meaningful for high-volume inference, where lower serving prices can change the economics of autonomous agents.
• –Developers should temper the hype: it lacks audio support, has weaker chat presentation, and can still choose poorly among many tools.