Google has made a significant move in the open-weight AI model landscape with the release of its Gemma 4 family, notably switching to the standard Apache 2.0 license. This is a game-changer for enterprises that previously shied away from Google's Gemma models due to their custom, restrictive licensing terms, which often involved usage limitations and the potential for Google to unilaterally update them. The new Apache 2.0 license removes these friction points, aligning Gemma with the terms offered by other leading open-weight models like Mistral and Qwen, making it a more attractive and legally straightforward option for commercial deployment and redistribution.
The Gemma 4 family comprises four models across two tiers: 'workstation' and 'edge'. The workstation tier features a 31B-parameter dense model and a 26B A4B Mixture-of-Experts (MoE) model, both capable of handling text and image inputs with substantial 256K-token context windows. The 'edge' tier includes the E2B and E4B models, compact units designed for mobile and embedded devices, offering text, image, and audio support with 128K-token context windows. The naming conventions highlight innovative parameterisation techniques: 'effective parameters' for the E-series models (like E2B, which runs like a 2B model despite having more total parameters) and 'active parameters' for the MoE model (like the 26B A4B, where only a fraction of parameters are active during inference), which significantly reduces computational costs.
Architecturally, Gemma 4 boasts native multimodality, integrating vision, audio, and function calling capabilities directly into its design, rather than treating them as add-ons. This means variable aspect-ratio image input, on-device audio processing with ASR and translation, and robust function calling are all built-in. The MoE architecture's use of 128 small experts, compared to the handful of large experts in other MoE models, is a key differentiator, promising efficient inference at a competitive performance level. Benchmarks show substantial improvements over previous Gemma versions, with the 31B model scoring impressively on reasoning and coding tasks. This release, combined with its permissive license and flexible deployment options (including serverless on Google Cloud), positions Gemma 4 as a compelling choice for organisations seeking advanced AI capabilities without licensing entanglements.
Fuente Original: https://venturebeat.com/technology/google-releases-gemma-4-under-apache-2-0-and-that-license-change-may-matter
Artículos relacionados de LaRebelión:
- Gemma 4 de Google IA Potente Licencia Libre
- Google Lanza Gemma 4 Modelos Abiertos Revolucionarios
- Arcees Trinity-Large-Thinking Americas Open Source AI Champion
- Cloudflare Unveils EmDash WordPresss Open-Source Successor
- Mozilla and Mila Champion Open Source AI
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario