July 2026 Summaries

Expedition Tiny Aya is paving the way for technology that adapts to human needs by utilizing open-source datasets, benchmarks, and methodologies to foster AI that bridges languages, cultures, and communities, thereby democratizing knowledge and opportunities globally. The initiative, which is part of the Cohere Labs Open Science Community, has gained traction with its core research being accepted to COLM 2026 and participation in projects like the Hugging Face Build Small Hackathon. Cohere Labs is a global network of researchers, students, and practitioners working collaboratively to advance AI research, offering programs, mentorship, and community events to encourage diverse contributions to AI's future. Madeline Smith, Operations and Community Manager at Cohere Labs, emphasizes the community-driven nature of this research and invites interested individuals to engage with their initiatives.

Speculative decoding (SD) is an innovative method designed to expedite large language model (LLM) inference by proposing multiple tokens with a smaller draft model, which are then verified by a larger target model, optimizing the use of GPU resources. Traditional SD faces challenges in production environments due to dynamic batch size (BS) changes, which can render it less effective when inference becomes compute-bound. Hardware-aware dynamic speculative decoding (DSD) enhances this process by adjusting the number of draft tokens (K) based on the interaction between the model and the hardware, increasing efficiency during memory bandwidth-bound scenarios and reducing K when compute-bound. This adaptability proves advantageous across various batch sizes and model architectures, as demonstrated in benchmarks comparing vanilla, fixed-K SD, and DSD configurations. Recent contributions to vLLM incorporate DSD, ensuring compatibility with async scheduling and full CUDA Graph, thus optimizing inference efficiency and maintaining the framework's performance in dynamic environments.

Cohere has launched Cohere Transcribe Arabic, an open-source automatic speech recognition (ASR) model designed to accurately convert spoken Arabic into text, excelling in capturing the nuances and dialectal richness of the language, which is spoken by over 300 million people across numerous dialects. The model demonstrates superior performance compared to leading alternatives like Whisper and OmniASR, achieving the lowest word error rate (WER) among open-source models and excelling in bilingual speech (Arabic-English) and varied acoustic conditions. Cohere Transcribe Arabic is optimized for enterprise use, providing high throughput and accuracy, and is available under the Apache 2.0 license, allowing developers to access and deploy it freely. This development addresses the gap in AI services for Arabic speakers, offering a solution that respects linguistic diversity and is adaptable to different professional contexts.

Cohere has developed a groundbreaking solution in the field of Automated Speech Recognition (ASR) with the launch of Cohere Transcribe Arabic, which is designed to bridge the gap in AI accessibility for Arabic-speaking users, who have historically been underserved compared to English-speaking markets. This model offers state-of-the-art accuracy by achieving the lowest word error rate on the Hugging Face Arabic ASR Leaderboard and excels in preserving dialectal nuances, code-switching, and maintaining enterprise terminology. Cohere Transcribe Arabic is optimized for high-throughput, making it suitable for production environments, and allows users sovereignty over their data and models by running efficiently on consumer hardware without external cloud dependencies. This open-source model is accessible through the Cohere API and Hugging Face, promoting democratized access to speech AI technology and encouraging developers to innovate and provide feedback through various platforms.