Inkling 2 Small lands on Hugging Face

This is a meaningful open-weight release, but “Small” remains a generous name for a checkpoint that demands serious infrastructure.

• –Sparse activation lowers per-token compute, but every expert still needs to be resident for inference.
• –NVFP4 makes the model more deployable on modern Blackwell hardware while keeping the checkpoint far beyond ordinary single-GPU use.
• –Hugging Face availability opens the door to self-hosted evaluation, inference-engine support, and community quantization.
• –Image inputs and long-horizon coding position Inkling 2 Small for agentic workflows, not just chat.
• –The real test is whether its 17B-active efficiency translates into competitive quality without frontier-model deployment costs.