Thinking Machines has rolled out Inkling-Small, a new open-source AI language model that packs a serious punch, rivaling its larger predecessor while being a mere quarter of its size. This move is a game-changer for businesses looking for powerful AI without the massive computational overhead. Inkling-Small is a multimodal reasoning model with a generous Apache 2.0 license, meaning it's flexible for commercial use.
What's truly impressive is that Inkling-Small, with its 276 billion parameters (12 billion active per token), nearly matches the performance of the 975 billion parameter Inkling. It achieves this by cleverly using a sparse Mixture-of-Experts architecture, activating only a fraction of its ex perts for each task. This efficiency translates to significantly reduced compute requirements, lower inference costs, and a smaller deployment footprint, making it more accessible for enterprises that may not have an abundance of GPUs.
While it doesn't quite fit on a laptop, Inkling-Small is far more manageable than its bigger sibling, making it ideal for businesses that can dedicate some GPU resources. It excels in areas like coding, reasoning, and multimodal tasks, even surpassing the original Inkling on several key benchmarks like SWE-bench Verified and Terminal Bench 2.1. However, for tasks demanding absolute factual accuracy, it still benefits from additional verification. Thinking Machines has made the model's weights readily available on Hugging Face and offers fine-tuning capabilities through its Tinker API. The company is also offering an attractive 50% discount for a limited time on API pricing, further lowering the barrier to entry for businesses wanting to leverage this advanced AI.
Fuente Original: https://venturebeat.com/technology/thinking-machines-debuts-inkling-small-open-source-ai-model-nearing-performance-of-predecessor-at-about-1-4-size
Artículos relacionados de LaRebelión:
- Thinking Machines Unveils Inkling Open-Source Multimodal AI
- AI Data Centres Schneider Foxconn Forge Power Pact
- Googles Gemma 4 12B Laptop AI Power
- GM Cars Get Google Gemini AI Power
- IBM AMD Chips Power Quantum Error Correction
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario