TypeSafe is an AI company that builds "System One" decision models, most notably Jev. Recent coverage centers on Jev: TypeSafe claims it is 20–200× faster and 40–400× cheaper on structured decision tasks and has opened early access. Developers have built on it, including the Jev Cards quiz game, while open projects such as Von and CUA-S1 cite Jev as inspiration. A researcher has accused TypeSafe of copying their research, and one analysis argues OpenAI may fast-follow the approach.
Developer launches picture quiz game built on Jev vision modelDeveloper Jose Sabater built Jev Cards, a picture quiz game around TypeSafe's new vision model Jev, which reads an image and answers in about a quarter of a second. The game offers solo leaderboards and live multiplayer, and its code is open source.
Cloudflare previews Clairic extraction modelCloudflare released a preview of Clairic, hosted on Workers AI as @cf/cloudflare/clairic, claiming higher accuracy, lower latency and lower cost than traditional LLMs for extracting, estimating and forecasting values from unstructured inputs. It is a System One extraction model, a category inspired by TypeSafe's Jev, and open weights are planned soon.
Researcher claims TypeSafe copied research as decision models flood marketA researcher has accused TypeSafe of copying their research, as lightweight "System One"-style decision models flood the AI market, with Microsoft and OpenAI among the companies releasing such models.
Home-Trained 0.8B Decision Model Jeff Runs in ~30msThe independent open project Jeff releases 0.8B and 2B zero-shot decision models fine-tuned from Qwen3.5 and Gemma 4, returning calibrated probabilities in a single forward pass at about 22 ms on an RTX PRO 6000 and 28 ms on an M4 Max, trained only on local hardware with synthetic data and released on Hugging Face.
Ollaya offers millisecond decision modelsOllaya is an open-source decision-model server for local hardware that answers in a single forward pass instead of token-by-token generation, handling a five-question request in about 10 ms. It exposes TypeSafe-compatible endpoints and works with the official Python SDK 0.7.1, shipping open models such as Laya, decider and nli.
Analysis: OpenAI may fast-follow TypeSafe's Jev modelAn Arcturus Labs piece argues TypeSafe's Jev model uses an approach close to conventional LLMs, so OpenAI is positioned to fast-follow and fold the capability into its models and agents. The key question is whether TypeSafe's training data and processes form a moat.
Kev: Tiny decision-model family built on Qwen3.5Developer jaredpalmer released Kev, a family of small decision models built on Qwen3.5, offered in 0.8B, 4B and 9B parameter sizes with pretrained weights, training code and evaluation data. The models handle yes/no, multiple-choice and rating questions, run on CUDA and Apple Silicon (4B and 9B fit a 32GB Mac in bf16), and expose an API matching TypeSafe System One.
Open-source 395M 'System One' model Von releasedA developer released Von, an open-source 395M parameter 'System One' model that serves as a drop-in replacement for TypeSafe's JEV. It runs entirely on CPU with 1–2 GB memory, responds in 25–300 ms, and beats JEV on all benchmarks. The developer is open to job offers from AI research labs.
CUA-S1 released: A System One model for computer useThe Cua team released CUA-S1-FORMS, a narrow decision model for form interactions with only 706k parameters and a 2.8 MB checkpoint, trained in under 30 minutes. Inspired by TypeSafe's JEV and Kahneman's dual-system theory, it predicts whether to use, check, click, or skip given elements, without generating new values or considering screenshots.
Reverse-engineered Jev-like model releasedAn open-source repo releases a Jev-like model that scores multiple text options in one pass, with Doom and chess examples. Reported results: Doom checkpoint averaged 0.60 kills and -97.50 reward over 10 episodes; chess checkpoint scored 4 wins, 46 draws, 0 losses vs a random mover but 0 wins, 2 draws, 48 losses vs Stockfish level 0.
TypeSafe launches Jev, a faster cheaper frontier modelTypeSafe announced its System One models and Jev, claiming Jev is 20–200× faster and 40–400× cheaper on structured decision tasks, using a new architecture, parallel sampler, and RLCD training; Jev is available in early access.
The A.I. Daily