Materne’s Software Catalyst: Thinking Through Small Language Models (SLMs) Role

研报英文原文证据摘录

Materne’s Software Catalyst: Thinking Through Small Language Models (SLMs) Role

Zoom: Zoom uses a federated AI architecture for Zoom AI Companion, dynamically

incorporating its own LLMs alongside third-party models including Meta Llama 2, OpenAI, and

Anthropic. The product applies different models to workplace tasks such as meeting summaries,

chat summaries, next steps, and response drafting.

Motorola: Motorola is embedding smaller, purpose-built language models into its Assist AI

platform to bring faster, more efficient intelligence closer to mission-critical workflows. Rather

than depending entirely on large cloud-based foundation models, MSI can use these smaller

models to reduce response times and operating costs, which is especially important in public

safety environments where dispatchers and first responders need accurate information in real

time. These capabilities support products across CommandCentral and Assist, including

automated call summaries, live transcription, and translation tools for 911 centers and personnel

in the field.

Microsoft Phi / Capacity: Capacity, an enterprise search company, uses Microsoft’s Phi small

language models through Azure AI Foundry for enterprise search workflows. The models support

offline enrichment tasks such as title generation and keyword tagging, as well as real-time query

refinement, with Microsoft citing 4.2x cost savings and 97% tagging accuracy.

Google Gemma: Google is positioning Gemma as an open-weight model family that developers

can tune and deploy across specific use cases, including summarization, extraction, translation,

and function calling. Recent variants include TranslateGemma, built on Gemma 3 in 4B, 12B,

and 27B sizes for translation across 55 languages, and FunctionGemma, a lightweight model