Inkling Small is an open-weight multimodal Mixture-of-Experts model from Thinking Machines Lab, featuring 276B total parameters with 12B active parameters per token. It natively accepts text, image, and audio inputs and generates text, supporting a context window of up to 524K tokens. Designed as a more efficient sibling to the flagship Inkling model, it delivers near-equivalent performance at roughly one-quarter the size. The model supports advanced reasoning with a configurable effort dial, and is intended for use in agentic systems, coding assistants, multimodal applications, and fine-tuning via Thinking Machines' Tinker platform.
Evaluations
8
across 8 benchmarks
Latency
1m 3s
Context length
524k
tokens
Cost
$0.58 · $1.44
input · output per 1M tokens
Key takeaways
Inkling Small is an open-weight multimodal Mixture-of-Experts model designed for generating text from text, image, and audio inputs. It aims to provide a more efficient yet high-performing alternative to the flagship Inkling model.
The model is a Mixture-of-Experts (MoE) architecture with 276 billion total parameters and 12 billion active parameters per token. It supports a large context window of up to 524K tokens.
Inkling Small demonstrates strong reasoning capabilities, consistently providing detailed step-by-step solutions for complex mathematical problems. Its limitations include minor formatting inconsistencies, occasional misinterpretation of symbols or units, and rare arithmetic errors.
The model achieved an impressive 98.8% accuracy on the MATH benchmark subset, excelling in algebra, number theory, geometry, and precalculus problems. It also scored 86.7% accuracy on the MMLU Pro benchmark.
Strengths on the MATH benchmark included consistently clear, step-by-step reasoning and strong performance in complex algebraic manipulation, number theory, and geometric calculations. Identified weaknesses were minor formatting issues (e.g., interval notation), misinterpretation of specific symbols (e.g., inequality, degree symbol), and occasionally missing roots in equations.
