AUS
NZ
UK
CA
USA
FRA
GER
ESP
ITA
NLD
PRT
SWE
GRC
IND
POL
BRA
ZAF
RUS
CHN
UAE
LONDON, UK - 29 January 2025 – Takara.ai, a trailblazer in Artificial Intelligence innovation, has announced the launch of SwarmFormer, a cutting-edge AI architecture inspired by natural swarm intelligence. This breakthrough delivers state-of-the-art performance while reducing computational resource demands by up to 94%, positioning the UK at the forefront of sustainable AI innovation.
This development aligns with the UK government’s AI Opportunities Action Plan, underscoring the nation’s ambition to lead in AI through ingenuity and practical solutions.
SwarmFormer: Nature-Inspired Efficiency
SwarmFormer mimics the collective behaviour of swarming insects to achieve remarkable efficiency. By combining local token interactions with cluster-based global attention, it matches the performance of industry-standard models while using only 6.7M parameters—a fraction of the 108M parameters in traditional systems. This innovation not only slashes infrastructure costs by 70%
but also enables advanced AI applications to run seamlessly on consumer-grade hardware.
“SwarmFormer is a game-changer,” said Jordan Legg, Chief AI Officer at Takara.ai. “It proves that the UK can lead in AI not just through investment but through innovation. By democratising access to powerful AI tools, we’re making it possible for organisations of all sizes to harness AI’s transformative potential.”
Pioneering Sustainable AI
SwarmFormer addresses key priorities in the government’s strategy, including:
Sustainable Infrastructure: Reducing energy consumption and computational overhead.Democratised Development: Lowering barriers to entry for smaller organisations.Homegrown Innovation: Establishing the UK as a global leader in AI research and application.
Technical Excellence with Real-World Impact
SwarmFormer leverages a hierarchical local-global attention mechanism, enabling decentralised multi-hop propagation of information and efficient global context representation. Its cluster-based architecture drastically reduces memory and computational requirements while retaining exceptional accuracy in text classification tasks. Experimental results show SwarmFormer achieving up to 90% fewer parameters than baseline models like BERT and outperforming them on key benchmarks.
For technical details, visit our SwarmFormer insights page. (https://takara.ai/thinking/insights/swarmformer/)
Distributed by https://pressat.co.uk/
