Cloudflare AI Capabilities Research

Yes. 81 open-source models on serverless GPUs across 300+ edge data centers. No GPU management. Pay-per-use.

LLMs: Llama 3.3 70B, Llama 4 Scout, DeepSeek-R1, Qwen3 30B, GPT-OSS-120B, Gemma 4 26B, Kimi K2.7, GLM-5.2, Mistral Small 3.1
Embeddings: BGE models, Qwen3, EmbeddingGemma
Speech: Whisper, Deepgram Nova-3, Aura TTS, Flux
Image: FLUX.2, SDXL, Leonardo Phoenix
Vision: Llama 3.2 Vision, Moondream 3, LLaVA
Other: Object detection, classification, content safety

2. Workers AI - Serverless Inference

Serverless GPU inference on Cloudflare edge
Free tier: 10,000 neurons/day
Paid: $0.01-0.50 per million tokens
No cold starts: <5ms via V8 isolates
Unified billing with AI Gateway (Aug 7 2026)

3. Custom Model Deployment

Capability	Support
LoRA fine-tuning	YES Llama, Mistral, Gemma
Custom models from scratch	NO
Model training	NO
Frontier models (GPT-4o, Claude)	VIA GATEWAY

4. Vectorize - Vector Database

Built-in vector DB for RAG, semantic search
Integrates with Workers AI embeddings
Up to 20M vectors per index
Metadata filtering

5. AI Gateway - Control Plane

Edge caching: <20ms, 0 tokens billed
Observability: requests, tokens, latency, costs
Fallbacks: auto-retry, route to backup models
Rate limiting: per-user/org/key spend limits
Universal API: swap providers without code changes
Unified billing: 5% surcharge or BYO-key free
Pricing: core features 100% free

6. Full AI Developer Stack

Layer	Product
Frontend	Pages
Compute	Workers (<5ms cold start)
AI Inference	Workers AI (81 models)
Orchestration	AI Gateway
Vector Search	Vectorize (20M vectors)
Stateful Agents	Durable Objects
Relational DB	D1 (SQLite at edge)
Object Storage	R2 (zero egress)
Background Jobs	Queues and Workflows
Key-Value	KV

7. Comparison: Cloudflare vs AWS vs GCP vs Azure

Feature	Cloudflare	AWS	GCP	Azure
Architecture	Edge (300+ PoPs)	Centralized	Centralized	Centralized
Cold Start	<5ms	500ms-30s	500ms-30s	500ms-30s
Model Catalog	81 curated	Hundreds	Hundreds	Hundreds
Model Training	No	Yes	Yes	Yes
Fine-Tuning	LoRA only	Full+PEFT+RLHF	Full+PEFT	Full+LoRA
Vector DB	Built-in	OpenSearch	Vertex Vector	Azure AI Search
Egress	$0	$0.08-0.12/GB	$0.08-0.12/GB	$0.08-0.12/GB
AI Gateway	Built-in free	Custom	Custom	Azure API Mgmt
Frontier Models	Via Gateway	Bedrock	Vertex	Azure OpenAI

Key Limitations

No frontier models on CF GPUs - route via Gateway
No model training - inference + LoRA only
Model size ceiling - ultra-large may not be on all nodes
Execution timeouts - Workers CPU ~5 min max
Vector ceiling - 20M vectors max

Bottom Line

Choose Cloudflare AI if: Building web/mobile apps, edge RAG, chatbots, AI proxies where low latency, zero cold starts, zero egress, and velocity matter.

Choose AWS/GCP/Azure if: Need model training, GPU clusters, exclusive frontier models, or billion-scale vectors.

For our Market Report App: Could run entirely on CF (Workers AI + Vectorize + KV + Pages). Hybrid: CF for pipeline + validation, external Claude for synthesis.

Cloudflare AI Capabilities

1. Does Cloudflare Host AI Models?