AI Models · sourc.dev

80AI Models

8Architectures

80LLMs

15Providers

80 AI models tracked across 8 architectures from 15 providers. LLMs account for the majority of tracked models and drive the largest share of ecosystem integrations.

80 AI Models entity_domains

8 Architectures subcategory

15 Providers providers

↳ Explore model architectures →

Architectures — by ecosystem weight

LLM 80

Large Language Models

Token-by-token language processing

VLM —

Vision-Language Models

Text and image understanding

SLM —

Small Language Models

Edge-optimized, low-latency inference

MoE —

Mixture of Experts

Sparse activation architecture

MLM —

Masked Language Models

Bidirectional context understanding

LAM —

Large Action Models

Language models that execute actions

SAM —

Segment Anything Models

Visual segmentation, no language component

LCM —

Large Concept Models

Sentence-level concept encoding

The concentration problem

models — GPT-4o and claude-3-5-sonnet — account for the majority of tracked software integrations across the entire ecosystem

entity_relations · 1952 active relations · 2026-04-09

What are AI model architectures?

Not all AI models are LLMs. The term LLM describes one specific architecture — models that process text token-by-token. sourc.dev tracks eight distinct architectures, each with different capabilities, use cases, and ecosystem roles.

What is the difference between an LLM and a VLM?

An LLM processes text only. A VLM accepts both text and image inputs. GPT-4o and Claude 3.5 Sonnet are both VLMs — they function as LLMs for text-only tasks and as VLMs when image inputs are provided.

Why track SLMs separately from LLMs?

SLMs run on-device rather than in the cloud. Pricing, capability, and integration patterns differ fundamentally from frontier LLMs. sourc.dev tracks 0 SLMs with verified benchmarks and deployment targets.

What qualifies as an AI model on sourc.dev?

Any model with verified API pricing, open weights documentation, or confirmed ecosystem integrations. Every entry has a signal_url and signal_date confirming the data source.

Data verified from primary sources. Methodology →