Cloudflare positions itself as a neolab with its own AI models

1 hour ago 11

The web infrastructure giant is now training small, fast models built to make decisions rather than write essays

Cloudflare has spent roughly a decade and a half as the plumbing behind a large share of the web. Now it wants a seat at a different table: it is building its own AI models.

At the AI Engineer Code Summit, Michelle Chen, Group Product Manager for Cloudflare’s AI Platform, described the company as a neolab. The term covers organizations that train models but are not named OpenAI or Anthropic.

Meet the models

The summit took place on November 11, 2026. By then Cloudflare had already shipped two models it calls System One models.

These are not chatbots meant to draft emails or debate philosophy. They are built for quick, efficient operational outputs, like sorting an input into a category or making a fast call, rather than open-ended generation.

The first pair, Clef and Clef-flash, went live on October 1, 2026. Clef has 27 billion parameters. Clef-flash has 9 billion.

Both models work with the Jev-API, support vision inputs, and handle context windows of up to 64,000 tokens.

In Cloudflare’s own testing, Clef-flash posted a median latency of around 39 ms. The company also reports that it matched benchmarks against Jev and other models on classification tasks.

Those are Cloudflare’s numbers, from Cloudflare’s tests. Independent benchmarks will tell a fuller story.

On October 19, 2026, Cloudflare added a third model called Clairic. It is an extraction model, designed to pull precise values like numbers and dates out of messy, unstructured inputs. Clairic is priced at $0.10 per million input tokens.

The weekend project that became a strategy

According to the research summarizing Chen’s talk, Cloudflare could not host the closed Jev model. The company then spent roughly one week deciding to train its own models instead.

Clef was trained over a weekend.

The company’s bet rests on infrastructure it already owns: distributed GPUs and edge computing. Edge computing means running workloads on servers physically close to users, instead of shipping every request to a distant central data center.

Chen’s demonstration showed how the pieces are meant to fit together. She combined Clairic with the Clef models and a larger LLM to run a diagnosis task.

The result was one call to the big model using about 1,800 tokens, compared with 40,000 tokens the conventional way. Cloudflare says that approach was approximately 3.5 times faster and 77% cheaper than relying on LLMs alone.

The logic is simple division of labor. Small models handle the triage and extraction. The expensive, heavyweight model only gets called when it actually needs to think.

What this means for the AI market

Cloudflare is not trying to out-OpenAI OpenAI. The current lineup targets a specific problem: the cost and latency of sending every small task to a giant model.

If Cloudflare’s demo numbers hold up in production, the pitch to enterprise customers writes itself. Cut token usage from 40,000 to about 1,800 on a task, and the savings compound quickly across millions of requests.

The neolab framing also signals ambition beyond today’s products. Cloudflare may develop reasoning models and other large language models down the line, according to the source material. That remains a possibility, not a roadmap.

Cloudflare already sits in the request path for a huge volume of web traffic. Offering cheap, fast models on that same network is a natural upsell to developers already on the platform.

Small models trained quickly can be brittle outside the tasks they were built for. Self-reported benchmarks are a starting point, not a verdict.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

Read Entire Article