Thinking Machines released Inkling, a 1-trillion-parameter open multimodal model, on Hugging Face on July 15, 2026. The system natively accepts image, text, and audio inputs across a 1-million-token context window.
Hugging Face published technical details showing that Inkling uses a Mixture-of-Experts decoder-only architecture containing 975 billion total parameters and 41 billion active parameters. The model incorporates 256 total experts, selecting six routed experts alongside two active shared experts per token. Instead of Rotary Position Embeddings, the model uses relative attention and alternates global attention layers with sliding window layers in a five-to-one pattern.
Model variants and deployment
The full bfloat16 checkpoint requires 2 terabytes of video memory, while an NVFP4 quantized variant requires 600 gigabytes. Recommended hardware setups include configurations such as 16 NVIDIA H200 GPUs or eight NVIDIA B300 GPUs. The release includes day-zero support for SGLang, vLLM, transformers, and llama.cpp, as well as multi-token prediction drafter layers for speculative decoding.
Thinking Machines Lab also launched Inkling-Small, a variant with 276 billion total parameters and 12 billion active parameters. Inkling-Small supports MXFP8 and NVFP4 formats and can be deployed to Hugging Face Inference Endpoints. On an eight-GPU RTX PRO 6000 setup costing $22 per hour, the smaller variant achieves up to 160 tokens per second. Unsloth provided 1-bit quantizations for running the architecture locally through llama.cpp.
Training and specialized features
Training for Inkling consumed 45 trillion tokens of text, image, audio, and video data. Vision inputs use a hierarchical multi-layer perceptron patchifier, while audio signals are processed through discretized mel spectrogram bins. For post-training workflows, Thinking Machines made available a managed tool called Tinker, providing reinforcement learning examples built with OpenEnv and the ECHO algorithm.
