Inkling Small

Thinking Machines Lab's official launch post introduces the Inkling family and explicitly previews Inkling-Small as a lighter-weight sibling with 12 billion active parameters, trained with a similar recipe to the flagship Inkling. The post confirms the model is available immediately for fine-tuning on the company's Tin The announcement positions Inkling-Small as a customization-friendly base model that achieves strong performance at lower cost and latency than the 975B-parameter flagship, which uses 41B active parameters with a 1M token context. Both models share native multimodal reasoning across text, images and audio with controll

VentureBeat reports that Thinking Machines released Inkling-Small, an open-source Apache 2.0–licensed multimodal reasoning model positioned as a smaller sibling to its flagship Inkling. Per the article, Inkling-Small has 276 billion total parameters with 12 billion active parameters per token, accepts text, image, and According to VentureBeat, Artificial Analysis assigned Inkling-Small a score of 40 on its Intelligence Index, within a single point of the larger flagship Inkling's score of 41 despite the smaller model being roughly one-quarter the size. The article frames Inkling-Small as a fit for enterprises with moderate GPU resou

Third-party coverage dated July 31, 2026 reports Thinking Machines launching Inkling-Small, a 276B-total-parameter MoE model with 12B parameters activated per token, achieving performance close to the larger Inkling. Full weights are released on Hugging Face alongside a Tinker Playground, and the model supports multimodal fine-tuning across text, image, and audio. The piece emphasizes that the open-weights release lets organizations self-host and fine-tune. The article details a variable thinking effort mechanism spanning minimal to xhigh, directly tied to output TFLOPs consumption for cost-quality tradeoffs. Benchmark framing positions Inkling-Small near Inkling on Terminal-Bench 2.1, HLE text reasoning, and IFBench instruction following. Training used NVIDIA GB300 NVL72 systems, and both original and NVFP4-quantized checkpoints are offered for Blackwell deployment, lowering local-inference barriers.

Thinking Machines Lab released an updated Hugging Face blog post confirming that Inkling-Small is now available as a companion to the full Inkling model. The post explicitly names Inkling-Small with 276 billion total parameters and 12 billion active parameters, sharing the same architecture as the larger Inkling, and a Day-0 runtime coverage is confirmed across transformers, SGLang, vLLM and llama.cpp for Inkling-Small, and a real-time voice and image demo is shipped so developers can interact with the variant directly. The post positions Inkling-Small as part of a broader collection of Inkling models and emphasizes its role as a lig

Thinking Machines Lab released Inkling-Small on July 30, 2026 under an Apache 2.0 license, publishing open weights and Tinker fine-tuning API access. The official model card describes it as a general-purpose multimodal model accepting text, image, and audio inputs and generating text outputs, intended for developers building coding assistants, agents, chatbots, and RAG systems. Distribution channels listed in the card include Hugging Face weight downloads and Tinker. The model is a 42-layer decoder-only sparse Mixture-of-Experts transformer with 276B total parameters and 12B active per token, routing each token to 6 of 256 experts plus 2 shared experts. Attention combines local and global layers, and a hierarchical patch encoder plus discrete audio token encoding handle images and audio. The card specifies BF16, MXFP8, and NVFP4 numerics, a 1M-token context window, 16kHz WAV audio input, and 40–4096px image input guidance for developers.