4 hours ago
- Inkling-Small is an efficient open-weights model that achieves comparable performance to Inkling at a quarter of its size.
- It is a Mixture-of-Experts transformer with 276B total parameters (12B active), trained on NVIDIA GB300 NVL72 systems.
- Features include native reasoning over audio and images, variable thinking effort, a 1M token context window, and strong benchmark performance.
- Inkling-Small surpasses Inkling on reasoning and agentic tasks while being more efficient, and matches or exceeds other models in its weight class.
- Full weights are released, and the model is available for fine-tuning on Tinker and chat via Tinker Playground.
- Improvements in training include a better pre-training data mix, post-training with on-policy distillation, and extended agentic coding RL.
- Multimodal capabilities are native, using an encoder-free architecture for audio and images, with strong performance on visual and audio benchmarks.
- Safety and epistemics are well-handled, with calibration, instruction following, and resistance to censorship matching Inkling's performance.