Inkling-Small Open-Weight Model Release by Thinking Machines
Thinking Machines released Inkling-Small, an open-weight model with 12 billion active parameters out of 276 billion total parameters. The model achieves comparable performance to the original Inkling at one-quarter of its size. It features native reasoning over audio and images and supports variable thinking effort. An NVFP4 checkpoint is available. Day 0 support is provided by vLLM. Using SGLang, the model achieves 648 tokens per second decode with DSpark (simulated accumulation length of 4) and 288 tokens per second without DSpark. The model is suitable for fine-tuning with NVIDIA NeMo on NVIDIA DGX Station.
Key insights
Inkling-Small Delivers Comparable Performance at Reduced Scale
Thinking Machines released Inkling-Small as an open-weight model featuring 12 billion active parameters within a 276 billion total parameter architecture. The model achieves performance comparable to the original Inkling model while operating at one-quarter of its size. This efficiency gain demonstrates significant optimization in model architecture, enabling broader accessibility for developers and researchers who may lack the infrastructure to run larger models. The release includes an NVFP4 checkpoint, further expanding deployment options.
Native Multimodal Reasoning and Variable Thinking Effort Expand Use Cases
Inkling-Small incorporates native reasoning capabilities over both audio and images, moving beyond text-only processing. The model also supports variable thinking effort, allowing users to adjust computational depth based on task requirements. These features position the model for diverse applications including fine-tuning workflows. NVIDIA highlighted its compatibility with NVIDIA NeMo on NVIDIA DGX Station, specifically recommending it for fine-tuning purposes.
Day 0 Ecosystem Support from vLLM and SGLang Ensures Immediate Usability
Inkling-Small launched with immediate inference support from vLLM on Day 0. Performance benchmarks using SGLang demonstrated 648 tokens per second decode speed with DSpark enabled at a simulated accumulation length of 4, and 288 tokens per second without DSpark. This rapid ecosystem integration reduces friction for developers adopting the model and indicates strong coordination between Thinking Machines and inference framework maintainers.