Hugging Face blog (Jul 15, 2026): day-0 welcome for Inkling.
- ~1T-param-class multimodal open model; natively image, text, audio inputs; 1M context
- Trained on 45T tokens; BF16 and calibrated NVFP4 variants; speculative MTP layers
- Day-0 support: transformers, SGLang, llama.cpp, vLLM mentioned
- Decoder-only multimodal MoE: 975B total / 41B active
- Focus: multimodal reasoning and domain adaptation via fine-tuning
- VRAM notes: very high for full precision; NVFP4 reduces footprint substantially