Several artificial intelligence companies announced new specialized models and capabilities this week, highlighting the growing trend toward domain-specific AI applications for enterprise use.

Copenhagen-based healthcare AI company Corti launched Symphony for Speech-to-Text, a clinical-grade speech recognition model designed specifically for medical environments. According to the company's research paper, the model achieved a 1.4% word error rate on English medical terminology, significantly outperforming general-purpose models including OpenAI's speech model at 17.7%, ElevenLabs at 18.1%, and Whisper at 17.4%. The model demonstrated particular strength in entity recall, achieving 98.3% accuracy on formatted clinical entities compared to 44.3% for general-purpose baselines. The system is designed to handle medical acronyms, medication dosages, and noisy emergency room environments that often challenge standard speech recognition systems.

Canadian AI lab Cohere released Command A+, a 218-billion-parameter language model under an Apache 2.0 open-source license. The model uses a Sparse Mixture-of-Experts architecture with only 25 billion parameters active during generation, designed to reduce computational requirements while maintaining performance. Cohere reported significant benchmark improvements, including jumping from 37% to 85% on τ²-Bench Telecom for complex reasoning and from 3% to 25% on Terminal-Bench Hard for coding tasks. The model supports 48 languages and includes native citation generation for enterprise compliance requirements.

Meanwhile, chip manufacturer Cerebras Systems announced it is serving Kimi K2.6, a trillion-parameter model developed by Beijing-based Moonshot AI, at nearly 1,000 tokens per second using its wafer-scale processors. Independent verification by Artificial Analysis confirmed speeds of 981 output tokens per second, which Cerebras claims is 6.7 times faster than the next-fastest GPU-based provider. The company's wafer-scale architecture stores model weights across multiple dinner plate-sized chips with 44 gigabytes of on-chip memory each, avoiding the interconnect bottlenecks that limit traditional GPU clusters.

Google also introduced Managed Agents in its Gemini API, promising to simplify agent deployment into a single API call while handling execution environments and tool integration. The service abstracts infrastructure complexity but requires developers to cede some control over the execution layer to Google's managed environments.

These announcements reflect a broader industry shift toward specialized AI models tailored for specific industries and use cases, as companies seek alternatives to general-purpose models that may not meet the precision requirements of regulated sectors like healthcare and finance.