Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarLatent Space (swyx)→ original

Cloudflare built a Jev equivalent in 2 weeks instead of 2 years and calls itself a neolab

Cloudflare has started calling itself a "neolab," the term for new-generation AI labs that build their own models. Its small model Clef makes fast decisions and replicates Jev from TypeSafe AI. Cloudflare built it in 2 weeks; the original took 2 years. The second model, Clairic, pulls an order number from an invoice in a single pass, without confusing a 3 with a 5, as LLMs sometimes do. Routine steps such as classification and field extraction are worth moving out of the large model.

AI-processed from Latent Space (swyx); edited by Hamidun News
Cloudflare built a Jev equivalent in 2 weeks instead of 2 years and calls itself a neolab
Source: Latent Space (swyx). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceLatent Space (swyx) ↗
Editor-in-Chief
Key metric #1Cloudflare's quarterly revenue is $782 million, up 39% year over year
No plans to build large LLMs from scratchthe team focuses on post-training
Key metric #3System One models respond in milliseconds, not seconds

For a decade and a half, Cloudflare sped up and protected other people's websites. Now it is releasing its own AI models and calling itself a "neolab." The term refers to a new generation of AI research labs that train their own models rather than only using someone else's.

The company's first two models belong to the System One class: small, fast models built for one narrow task, which respond in milliseconds, not seconds. Clef makes decisions and follows the idea behind Jev, a model from the startup TypeSafe AI. Cloudflare built it in 2 weeks, while the original creators spent 2 years on Jev.

Clairic extracts values from unstructured text, such as an order number from an invoice. The practical takeaway for developers: not every step in a pipeline needs to go through a huge LLM.

What the models are and where System One fits in

The idea behind System One is to avoid sending every request to a large language model. When you need a quick decision, such as routing a request to the right destination, a small specialized model is cheaper and more predictable. Michelle Chen, who leads product for Cloudflare's AI platform, said as much at the AI Engineer Code Summit: "We don't need to send every request to an LLM. Some of it can go to a System One model."

Clef is exactly this kind of decision model. According to Chen, Jev became an "east wind" for the company in September, unexpectedly changing how it thinks about models. She explains the gap between two weeks and two years by noting that the team already knew what the API schema should look like. She gives all the credit to TypeSafe AI, which invented this new type of model and brought it to a working state.

Clairic is Cloudflare's own development, launched a month ago as the first model in a new class of "extraction models." It takes unstructured input and extracts, computes, estimates, or predicts values. Chen describes the problem: LLMs generate text one token at a time, so they can hallucinate and sometimes return broken JSON. Take an invoice: a large model suddenly doubts whether the order number is three, four, or five, even though it is written plainly in the document. Clairic is meant to return a typed value, an actual number, in a single forward pass.

Neolab or marketing

Stephanie Palazzolo of The Information proposed the term "neolab" last November. Diddy Das of Menlo Ventures lists Thinking Machines Lab, Poolside, and TypeSafe AI itself. According to Das, neolabs employ elite researchers and spend heavily on compute (usually to train their own models). They have no notable revenue and are typically startups with large funding rounds.

Cloudflare fits this description at best halfway. It is a public company with substantial revenue: $782 million in the third quarter, up 39% year over year, with annual revenue topping $3 billion. The author of the piece admits the label partly sounds like marketing. Chen frames it as a natural evolution: infrastructure, compute (the Workers platform, now for agents too), storage, running the open models Kimi, GLM, and DeepSeek in Workers AI, and now its own models.

Why the company wants its own models

Chen gives two reasons. The first: "Being a neolab means building what we are missing." The second is risk reduction. Jev is closed, and you cannot run it yourself. Chinese open models may carry regulatory risks. And many labs want to make money and are introducing expensive licenses.

The plans are modest, though. Cloudflare does not intend to pretrain large models from scratch, as OpenAI and Anthropic do. In Chen's words, "most of our team works on post-training." Reasoning models and other LLMs are possible, but built on existing models made more efficient. For now these are hints, not announcements. The talk included no quality metrics for Clef or Clairic, so it is not possible to compare their accuracy with large LLMs.

What to do in practice

Go through your pipeline and mark the steps where an LLM handles classification, routing, or pulling fields out of invoices, emails, and logs. These are candidates for a small specialized model: responses are faster, the output format is more stable, and there is less risk of an invented order number. For Cloudflare, it is also a way to sell more services on its own infrastructure, where millisecond latency is the main competitive advantage.

Story

Key Questions Answered

What the models are and where System One fits in?

The idea behind System One is to avoid sending every request to a large language model. When you need a quick decision, such as routing a request to the right destination, a small specialized model is cheaper and more predictable. Michelle Chen, who leads product for Cloudflare's AI platform, said as much at the AI Engineer Code Summit: "We don't need to send every request to an LLM. Some of it can go to a System One model." Clef is exactly this kind of decision model.

What are the key details on Neolab or marketing?

Stephanie Palazzolo of The Information proposed the term "neolab" last November. Diddy Das of Menlo Ventures lists Thinking Machines Lab, Poolside, and TypeSafe AI itself. According to Das, neolabs employ elite researchers and spend heavily on compute (usually to train their own models). They have no notable revenue and are typically startups with large funding rounds. Cloudflare fits this description at best halfway.

Why the company wants its own models?

Chen gives two reasons. The first: "Being a neolab means building what we are missing." The second is risk reduction. Jev is closed, and you cannot run it yourself. Chinese open models may carry regulatory risks. And many labs want to make money and are introducing expensive licenses. The plans are modest, though. Cloudflare does not intend to pretrain large models from scratch, as OpenAI and Anthropic do.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…