Cloudflare launches Clef-omni multimodal model, cuts hosted Clef-flash context to 24k

Cloudflare's Clef family expands with Clef-omni, a fully multimodal model, plus speed and price updates to the existing Clef and Clef-flash tiers. The announcement positions Cloudflare as a model provider in its own right, not just an inference host on its edge network.

The controversial detail is context. To reach a lower price point, Cloudflare reduced hosted Clef-flash's context window to 24k tokens, down from the 64k it had previously advertised. The open weights on Hugging Face still support 256k tokens, meaning developers who self-host get roughly ten times the context available on Cloudflare's own service.

The trade-off reflects inference economics: long-context serving is memory-hungry, and capping context lets providers pack more requests per GPU. Cloudflare is effectively segmenting — cheap, short-context hosted calls versus full-capability self-hosting — a pattern similar to providers restricting features of open models on hosted endpoints.

Developers have been vocal about the downgrade, particularly those who built against the advertised 64k. The move also contrasts with competitors racing to longer contexts, such as xAI's 500K-token Grok 4.7 and Mistral Large 4's 1M. Watch whether Cloudflare offers a paid long-context tier for Clef-flash or reverses course.