---
title: "Industry Insight | Multimodal and Smaller Models Are Replacing Giant LLMs as Many Enterprises’ First Choice"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2024-05-28T18:00:00+08:00"
canonical: "https://ffcap.cn/en/research/20240528a07wxj00"
source: "https://ffcap.cn/en/research/20240528a07wxj00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20240528A07WXJ00"
---

# Industry Insight | Multimodal and Smaller Models Are Replacing Giant LLMs as Many Enterprises’ First Choice

_Unique Research · Insight · May 28, 2024_

In the GenAI boom, enterprise IT leaders are looking past pure LLMs. Not every business scene fits a large language model. Emerging **multimodal** models and **smaller** models are showing more purpose-fit potential.

LLMs still define GenAI in most experiments — coding assistants, text-to-image stacks that pair LLMs with diffusion. But some CIOs say LLMs aren’t always the right tool; they’re exploring multimodal systems that handle dynamic tabular data in spreadsheets and vector DBs, plus video and audio.

IDC’s _Generative Foundation AI Model Market Overview_: multimodal foundations combine text, audio, image, video — caption images or answer visual questions. Examples: Google Gato, OpenAI GPT-4o, Microsoft LLaVA, Nvidia NeVA, Vicuna, BLIP2, Flamingo.

## Healthcare multimodal — Northwestern Medicine × Dell

Northwestern Medicine’s advanced-tech group with Dell AI Innovation built a proprietary multimodal LLM that reads chest X-rays and summarizes findings — patients get results ~**80%** faster. Next: CAT/MRI multimodal and full-EHR predictive models. Dr. Mozziyar Etemadi: radiologists already save ~**40%** of text-dictation time, plus more from image analysis — “we used to be stuck on LLMs, text, or Excel; now we handle images and X-rays. A huge leap.”

## Beyond LLMs — MakeShift’s large graph models

Workforce-scheduling SaaS MakeShift (healthcare, retail, manufacturing) found LLMs fine for admin/staff chat but inadequate for vector data and huge interrelated graphs when optimizing forecasts. CTO Danny McGuinness: they switched to a patent-pending **Large Graph Model (LGM)** from MIT spinout **Ikigai Labs** — no-code models for tabular/time-series, with Medico, HSBC, Spirit Halloween, Taager.com, Future Metals, WIO. Ikigai (Devavrat Shah & Vinayak Ramesh) doubled headcount in six months; raised **$25M** late last year.

Scheduling is 24/7 plus union rules. Since last year MakeShift built on Ikigai APIs — now in production. Benefits: predictive scheduling on shifting data; customers adding weather, transit distance, in-store density; healthcare clients exploring finance-process use cases beyond historical schedules. New correlations/causal links surface that teams hadn’t noticed.

Databricks AI VP Naveen Rao notes LLMs can also handle tabular and other forms via markup — relevant after Databricks acquired his company Mosaic. With LGMs and LLM limits, GenAI competition is fierce; CIOs face crowded, uncertain choices.

## Right-sized small models

Gartner’s Arun Chandrasekaran: multimodal evolution was expected, but costly compute/data keep commercial share modest. 2023 was mostly text/code; then vision and early speech. “Cost of compute and data to build these is still very high.” Hence many enterprises move past giant LLMs to smaller models that are “good enough” and cheaper — market will periodically reprice model size.

Rao: building an LLM can cost up to **~$200M** — mostly labeling/data management, not just FLOPs. Mosaic’s thesis: affordable, usable models; specialization beats generalization for most firms. Path: train from scratch on your data, or fine-tune open models. Baldor Specialty Foods CIDO Satyan Parameswaran (ex-UPS): prefer small models to avoid LLM hallucinations — grab one from Hugging Face and customize for the task.

## A new GenAI route

Vendors shipping smaller models include C3.ai, Anaplan, Dataiku, Hugging Face. Ikigai CEO Shah: LGMs give probabilistic representations of tabular timestamp data — learn variable relations, fill missing data, match similar rows across sheets, detect change points/anomalies, simulate/synthesize duration _on your data only_.

Cost still gates customization. Text-only LLMs already burn huge compute; as chipmakers and clouds race silicon, enterprises will mix large and small models in production. Many still start with LLM experiments then harden. LVM/LGM use is early, but MakeShift-class adopters already see value: put the right skills in the right place at the right time — seniority, site rotation, union rules, burnout, overtime — complexity that is crushing without AI, and much easier with multimodal and task-focused small models.

_Author: Paula Rooney · Compiled: Qiuping. Charts attach on ingest (5 images)._

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240528a07wxj00
On-site reading page: https://ffcap.cn/en/research/20240528a07wxj00