Hugging Face Blog
·
1 month ago
JetBrains released Mellum2, a 12-billion-parameter Mixture-of-Experts model trained on natural language and code under the Apache 2.0 license. The model activates only 2.5 billion parameters per token while delivering more than twice the inference speed of similarly sized competitors. Mellum2 is designed for latency-sensitive tasks like routing, retrieval-augmented generation, summarization, and agent subtasks within larger AI systems.
Hugging Face Blog
·
1 month ago
IBM implemented agent logic—software primitives like knowledge graphs and program analysis libraries—to guide large language models through enterprise workflows in four domains: legacy code understanding, test generation, incident response, and compliance automation. The approach achieved up to 30× lower token consumption in legacy code analysis, 20-45% improvements in test coverage with 15× fewer tokens, 4.0× better incident investigation performance, and 1.3-2.0× better compliance outcomes compared to baseline LLM-only or competing agent approaches. By constraining LLM reasoning through structured task decomposition and domain-specific logic, enterprises can reduce costs while improving accuracy and adoption of AI systems in mission-critical workflows.
