thinkingmachines/Inkling-Small
3 hours ago
- Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
- It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
- The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.
- Training data is sourced from public, third-party, and synthetic data, with cleaning and filtering to improve quality and safety.
- Evaluations show strong performance across agentic, reasoning, factuality, vision, audio, and safety benchmarks.
- Safety mitigations include red-teaming and refusal suppression testing; residual risks like compliance with harmful role-play are addressed with defense-in-depth and moderation tools.
- Limitations include hallucination, biases, uneven multilingual performance, and knowledge cutoff; human oversight is recommended for high-stakes applications.
- The model is released with open weights for research, fine-tuning, and integration into third-party products.