🧠Inkling 2 Small Ships With 594B Parameters and Open Apache 2.0 Weights Next Week

Your coding agent just got a serious upgrade

Inkling 2 Small is here, and it's a real step up from the first generation. It's a Mixture-of-Experts model with 594B total parameters and only 17B active, so you get big-model capacity without big-model inference costs. It accepts image inputs and variable effort, and it scores 31 on the Artificial Analysis Intelligence Index, far ahead of previous models. If you're building agents, this one matters. It was trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses, so it's built for loops that run for hours, not single prompts. It follows instructions better, checks its own work without being asked, and uses visual feedback to refine websites, games, and 3D scenes. The numbers are where it gets interesting. It hit 87.7% on Vibe Code Bench v1.18 and 82.4% on SWEBench Pro v2 (Hard). On kernel optimization it posted a 6.02x preliminary speedup. Initial access is through Tinker with closed weights, and the Tinker budget for a Qwen3.5-4B LoRA run was $426. The catch? Weights aren't out yet. They ship next week, after external safety testing and red-teaming, so wait for the checkpoints before you plan a self-hosted rollout.