1. Cloudflare AI models mark a change of identity
Cloudflare, the web infrastructure company, has begun training its own artificial intelligence models and now describes itself as a "neolab." It released two decision models, Clef and Clef-flash, on October 1, 2026, and an extraction model called Clairic on October 19, 2026, with a product manager presenting the strategy at the AI Engineer Code Summit in November.
The label was used by Michelle Chen, a product manager on Cloudflare's AI Platform. In industry usage, a neolab is a company that trains models of its own without being one of the established frontier developers such as OpenAI or Anthropic. Applying the term to a business that has spent roughly 15 years known for content delivery, security and network services is a statement about where it believes its future lies.
The Cloudflare AI models are not attempts to build a rival general-purpose chatbot. They are narrow tools designed to do small jobs very quickly, and the company's argument is that a great deal of what developers currently send to large, expensive models could be handled by something far lighter.
2. What is Cloudflare Clef?
Clef is a 27-billion-parameter model that Cloudflare built for fast operational decisions, such as classifying an input or choosing between a small set of outcomes. It launched on October 1, 2026, alongside a lighter sibling, Clef-flash, which has 9 billion parameters.
Cloudflare groups both under the heading of "System One" models. The phrase borrows from the idea of two modes of thinking, one quick and automatic and the other slow and deliberate. A System One model, in this framing, is the quick mode: it does not reason through a problem at length or produce paragraphs of text, but returns a rapid judgment on a tightly defined question.
Both models work with the Jev API, the interface of an existing decision model, which means developers already using that format can direct their requests to Clef without rewriting their applications. Both also accept vision inputs, so they can make decisions about images as well as text.
Parameter counts give a rough sense of scale. Parameters are the numerical values a model adjusts during training, and more of them generally means greater capacity along with greater cost to run. At 27 billion and 9 billion, the two Clef models are small next to the largest systems on the market, and that modest size is deliberate: a model that only has to choose among a few options does not need the breadth of one expected to write essays or computer code.
Speed is the headline claim. In Cloudflare's own tests, Clef-flash posted a median latency of about 39 milliseconds and performed on par with Jev and other models on classification tasks. Those numbers come from the company itself, and independent benchmarks will be needed before they can be treated as settled.
3. Clairic, the extraction model
The third release arrived less than three weeks after the first two. Clairic, added on October 19, 2026, is an extraction model: its job is to pull specific values, such as numbers and dates, out of unstructured text.
Extraction is one of the most common chores in business software. Invoices, contracts, support tickets, medical notes and emails all contain facts that need to be lifted out and placed into structured fields before another system can act on them. Developers frequently hand that task to a large language model, which works but means paying for a general-purpose system to perform a narrow, repetitive function.
A purpose-built extractor changes that equation. Because it is asked only to locate and return particular values, it can be smaller and quicker than a system designed to hold open-ended conversations, and its output is easier for downstream software to consume. In a pipeline, it acts as the first filter, turning a long and messy document into a handful of clean fields that later steps can work with.
Cloudflare has priced Clairic at $0.10 per million input tokens. Tokens are the units of text that models process, and pricing by the million is the standard way AI services are billed. A low price on a specialized model is central to the pitch, because extraction workloads tend to involve very high volumes of documents.
4. How a model Cloudflare could not host led to a new strategy
The origin of the effort, as recounted in summaries of Chen's talk, was a practical obstacle. Cloudflare wanted to offer the Jev model on its platform but could not host it, because the model is closed.
The company took about a week to decide that the answer was to train models itself. What followed was fast by any standard: Clef was trained over a weekend.
That timeline says something about the kind of model involved. Training a frontier-scale system is a project measured in months and enormous computing budgets. Training a compact model for a constrained task is a different undertaking, and one that a company with ready access to graphics processors can attempt in days.
Cloudflare's bet rests on hardware it already operates. The company runs a distributed fleet of GPUs as part of its edge computing network. Edge computing means running workloads on servers located near the people using them, instead of in a small number of distant central data centers. For models whose entire selling point is responding in tens of milliseconds, physical proximity to the user is an advantage that is difficult for a centralized provider to replicate.
5. The demo: one large-model call instead of many
Chen illustrated the approach with a diagnosis task that chained the new models together with a larger language model. Clairic handled extraction, the Clef models handled classification and routing, and the large model was brought in only for the step that required it.
The effect on consumption was dramatic. The combined pipeline made a single call to the large model and used about 1,800 tokens, compared with 40,000 tokens when the task was run the conventional way. Cloudflare says the cascade was about 3.5 times faster and 77% cheaper than relying on large language models alone.
Token counts matter because they drive both the bill and the wait. Large language models charge according to how much text they read and write, and the time they take to answer grows with the length of what they must process. A task that feeds tens of thousands of tokens into a large model is therefore expensive and slow by construction, whatever the quality of the result.
The principle is a division of labor. Small models perform triage and pull out the relevant facts, and the expensive model is consulted only when its broader capabilities are needed. It mirrors how many organizations already structure human work, with routine sorting handled quickly at the front and specialists reserved for the cases that warrant them.
For companies running AI features at scale, the arithmetic is what matters. If the demonstration's results hold up in production, reducing a task from 40,000 tokens to roughly 1,800 would translate into substantial savings once multiplied across millions of requests, along with noticeably faster responses for end users.
6. Where this leaves Cloudflare in the AI market
None of this puts Cloudflare in direct competition with the largest model developers on their own terms. Its lineup targets a specific inefficiency: the cost and delay incurred when small, simple tasks are routed to giant models that are far more capable than the job requires.
That gap has grown as AI features have moved from experiments into everyday products. Early adopters often sent every request to the most capable model available, because simplicity mattered more than cost while volumes were low. As usage climbs, the same habit becomes a significant line item, and engineering teams start asking which steps truly need a frontier model and which could be served by something smaller. Cloudflare is positioning its lineup as the answer to that second question.
Distribution is the company's other asset. Cloudflare already carries a large share of the web's traffic, and a sizable developer community builds applications on its network. Offering cheap, fast models on the same infrastructure those developers already use is a natural extension, since it adds a capability without asking customers to adopt a new vendor.
The topic has relevance well beyond web hosting. Automated agents that execute transactions, monitor markets or screen activity for compliance depend on rapid, inexpensive classification and extraction steps, and the cost of those steps shapes what kinds of automation are economical. Cheaper building blocks widen the range of tasks that can be handed to software.
The neolab framing also hints at broader ambitions. Describing itself in those terms leaves room for Cloudflare to pursue reasoning models or larger language models in the future. That is a possibility implied by the positioning, however, and not a roadmap the company has announced.
7. Reasons for caution
Two caveats temper the picture. The performance figures, from the 39-millisecond latency to the 77% cost reduction, are Cloudflare's own measurements, and self-reported benchmarks are best treated as a starting point until outside testers reproduce them.
The second concern is robustness. Small models that are trained quickly can prove brittle when they are pushed beyond the tasks they were built for. A classifier that performs well on the inputs it was tuned on may behave unpredictably on unfamiliar ones, and developers adopting these tools will need to test them against their own data before depending on them.
Whether the approach spreads will depend on how those two questions are answered. Independent evaluations of Clef, Clef-flash and Clairic, and evidence from customers running the cascade in live applications, will show whether the gains seen on stage survive contact with real workloads.
