AI Signal 242 2 feeds carried it
Inkling Small from Thinking Machines is now available on AI Gateway
↗
Engineers gain access to a natively multimodal model (text, image, audio) that only activates 12B parameters at inference time, significantly reducing compute overhead compared to dense models of similar capability. Day-0 integration with vLLM, SGLang, llama.cpp, and Hugging Face Inference Endpoints means deployment paths are already established rather than waiting on community implementation.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Inkling-Small is a Mixture-of-Experts model with 276B total parameters but only 12B active per inference pass, and supports quantized weight formats NVFP4 and MXFP8 for further efficiency.
The architecture diverges from standard transformers by using relative attention instead of RoPE, hybrid sliding-window-to-global attention at a 5:1 ratio, and short 1D convolution (SConv) for local context.
Both Hugging Face and Vercel are providing immediate deployment options, with Hugging Face offering one-click Inference Endpoints achieving up to 160 TPS.
THE CLUSTER