We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·21d ago·5 min read
Welcome Inkling by Thinking Machines
TL;DR
Inkling by Thinking Machines is a groundbreaking multimodal large language model (LLM) that processes image, audio, and text inputs with a 1 million context window. It features a unique architecture with hybrid attention and relative attention mechanisms, enabling efficient multimodal reasoning.
✦ Why It Matters
Engineers can start building multimodal applications today using Inkling's architecture and capabilities for enhanced reasoning tasks.
Key Takeaways
- Utilize Inkling for multimodal reasoning tasks.
- Explore fine-tuning options for domain-specific applications.
- Leverage the model's efficient inference capabilities.
- Deploy Inkling using SGLang or vLLM for optimal performance.
How It Works
Inkling's architecture is based on a decoder-only model that supports causal autoregressive generation. It employs a Mixture-of-Experts (MoE) strategy, activating only a fraction of its parameters during inference to enhance speed.
The model uses relative attention to encode positional information directly in the attention logits, and hybrid attention layers alternate between global and sliding window attention for efficient computation. This design allows Inkling to effectively process and reason across text, images, and audio inputs.