TypeSafe's $40M Model Launch Ran Into a Researcher's Year-Old Paper

TypeSafe AI came out of stealth on Sept. 15 with $40 million in seed funding led by DCVC and a model called Jev that makes structured decisions instead of writing text. Four days later, an independent researcher published a paper trail arguing he had built the same architecture more than a year earlier.

TypeSafe describes Jev as a "System One model," a category the company says takes unstructured program state as input and returns typed, probabilistic decisions in a single parallel pass rather than generating text one token at a time. Because valid outputs are constrained to a predefined schema, TypeSafe says the model cannot produce a type error. Diogo Almeida, TypeSafe's co-founder and chief executive and a former OpenAI researcher credited as a co-inventor of reinforcement learning from human feedback, told The Register: "I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can't be the only consumers of intelligence."

What Jev costs and how fast TypeSafe says it runs

Jev prices input tokens at $0.042 per million and does not charge for output. The Register, which viewed a demo, reported the model answering in 0.114 seconds against 8.566 seconds for a comparison model it called GPT-5.6 Terra, and TypeSafe's own materials put end-to-end latency at 70 to 500 milliseconds, calling that 40 to 200 times faster than frontier language models on comparable tasks. Access is API-only through TypeSafe's console; the company has not published a technical paper describing Jev's architecture or released its weights.

The papers Mukkunnoth says came first

Nandakishor Mukkunnoth, founder and chief executive of ConvAI Innovations, wrote on Dev.to that TypeSafe had "proposed the exact same non-autoregressive decision concept as if it was a brand-new scientific breakthrough." He points to two of his own papers: one posted to arXiv on March 30, 2025, describing a reinforcement-learning model that predicts sales-conversion outcomes in 85 milliseconds using frozen sequence embeddings and a separate value network, and a second posted Sept. 23, 2025, laying out schema-based decisions guided by reinforcement learning. Mukkunnoth wrote that TypeSafe launched Jev "without technical papers, without open weights, and with zero open training datasets," a description consistent with what TypeSafe has published so far.

Laya, the answer he built instead of a dispute

Rather than pursue a credit dispute, Mukkunnoth released an open-source model called Laya through ConvAI Innovations, licensed under Apache 2.0. Laya returns calibrated probabilities across three response types, choice, score and a yes-or-no primitive, and Mukkunnoth says it runs in 32.8 milliseconds on a single GPU, or 7.2 milliseconds per question when batched, which he describes as six to eight times faster than Jev's published 150-millisecond latency. Mukkunnoth wrote that the experience of watching a funded lab attract attention for a concept he had shipped as open-source software for a narrower use case was "incredibly frustrating."

Model	Latency	Access
TypeSafe Jev	~150 ms published	API only, closed weights
Mukkunnoth's Laya	32.8 ms (7.2 ms batched)	Open weights, Apache 2.0

TypeSafe has not addressed Mukkunnoth's papers or Laya publicly, and its Sept. 15 announcement predates his post by four days. The Register's own coverage, published the day of the launch, raised a separate objection: that a "hallucination-free" structured output is not the same as a correct one, since a valid, schema-conforming answer can still be wrong. Between the two critiques, the open question TypeSafe has yet to answer is not whether Jev is fast, but whether the architecture behind it is the novel one the company's launch described. It is the same question this month's other vendor-benchmark disputes keep raising, from OpenAI's own chip claims to a checkpoint bug in Cua's CUA-S1-Forms that surfaced only after launch. TypeSafe's next scheduled move, a wider release beyond its current console-only access, has no announced date.