TypeSafe's Jev Raises $870M at $7.5B Valuation: Decisions, Not Chat

The Funding

On October 9, Bloomberg reported that TypeSafe AI closed roughly $870 million in funding at a $7.5 billion valuation. The round was led by a16z, with Sequoia and existing investor DCVC participating.

Three numbers stand out.

First, speed. Jev only officially launched on September 15 — 24 days from launch to a closed round.

Second, the jump. On September 15, TypeSafe also announced a $40 million seed round (led by DCVC). In under a month, the raise went from $40 million to about $870 million.

Third, versus the rumors. The Information had reported over $1 billion raised at a $10B+ valuation — the closed numbers came in below that.

What Is Jev

Jev also runs on a transformer architecture. It generates no text — its outputs are probabilities, choices, or scores. The company calls these “calibrated decisions,” trained via RLCD (Reinforcement Learning for Calibrated Decisions). Internally it’s dubbed a System One model, borrowing psychology’s “fast thinking.”

Jev supports only three request types. First, answer yes/no questions with a probability, called Noul. Second, pick one item from a given list. Third, score against a rubric — say, rating a security alert’s severity or estimating a support ticket’s urgency.

Founder Diogo Almeida worked on InstructGPT and RLHF, spending 2018–2024 at OpenAI after Google Brain. In 2024 he left OpenAI to found TypeSafe with former Meta research engineer Sasha Sheng and engineer Erik Gafni.

His argument is blunt: computers don’t speak human. Software needs a stable, repeatable output layer — not text that must be parsed again. Frame the question well, let the model only judge, and skip the parsing and rewriting code.

Why Enterprises Pay

By TypeSafe’s own account, about a third of Fortune 500 companies use Jev. Almeida told the WSJ it’s about 25%, processing roughly a trillion tokens a day. Both figures come from the company; neither is independently verified.

Many software tasks end in a judgment: approve or reject, workflow A or B. Doing that with an LLM used to mean generating text, then converting it to structured data — an extra parsing layer and an extra chance to fail. Jev returns a typed answer directly.

On October 1, Cloudflare released open-weight Clef and Clef-flash, API-compatible with Jev. The WSJ called it one of the clones.

Two Unanswered Questions

First, calibration quality. Jev’s core selling point is honest probabilities. As TechCrunch notes, there’s no independent verification of how well these decision models are actually calibrated in the real world.

Second, “no hallucinations.” Jev outputs no natural language, so it can’t confidently spout nonsense. But it can still pick the wrong option. Not hallucinating isn’t the same as not erring.

What to Watch

The company hasn’t detailed how the money gets spent. Two things to watch: whether the Fortune 500 usage numbers get independently verified, and whether Jev’s probability calibration survives auditing in production. The open-source clones are already catching up.