Frontier AI & Model Developers

TypeSafe AI's 'Jev' bets on a non-LLM architecture — and draws $10B+ interest

Source: Bloomberg / Financial Times · Sep 25, 2026

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, is building a model called 'Jev' that deliberately isn't a language model. Instead of predicting the next token, it outputs decisions directly as probabilities and confidence scores, and the company claims it can be roughly a hundred times faster and cheaper than an LLM for certain classification and scoring tasks. After a $40 million seed from DCVC in mid-September, the Financial Times reports funding offers valuing the company north of $10 billion — a startling jump from an earlier mark near $200 million.

Whether or not the specific numbers hold, the architecture is interesting because of where it aims: at judgment and scoring, not conversation. Models built to output calibrated decisions need training data that captures how real experts actually judged something — the input, the reasoning, the verdict, and ideally the outcome that followed.

That is the exact shape of a structured interview record: a human evaluator's confidence-scored assessment of a real response. As more of the field explores decision-native models rather than text-native ones, the demand shifts toward supervised signal about human judgment — which is far harder to synthesize than fluent text, and only exists where real evaluations were captured with provenance intact.

Key Points

• Ex-OpenAI researcher Diogo Almeida's model outputs decisions as probabilities/confidence scores rather than predicting tokens
• Claimed ~100x faster/cheaper than LLMs on certain classification and scoring tasks
• $40M seed (DCVC) in mid-September; the FT reports funding offers valuing it at $10B+, up from ~$200M
• A genuinely different architecture aimed squarely at decision/scoring workloads — the layer where evaluation data matters most