Amazon releases its own Jev clone as decision models flood the web

Releases·Model Releasefeatured

TypeSafe AI launches Jev, an RLCD-trained decision model claiming 20-200x faster, 40-400x cheaper inference than LLMs

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.

Researchers released this-that-model-1.0, a 2B-parameter typed decision model that reads answers directly from hidden states, outputs one of n declared options plus a thresholdable number, and generates zero tokens. It decides in 30.9 ms on one laptop GPU versus 8758 ms for a frontier API call, and scores 0.941 (Brier 0.042) on a third party's 68-question cohort versus 0.765 (Brier 0.133) for hosted service Jev. It is open-sourced on Hugging Face, targeting software branches that need decisions, not prose.

Releases·Model Release

Trained locally: ultra-fast 0.8B/2B System 1 decision models that match Jev on benchmarks and Doom, ~30 ms per decision (open weights)

A developer fine-tuned Qwen3.5 0.8B/2B and Gemma models into zero-shot classifiers ('Jeff models') that return calibrated probabilities in one forward pass, releasing Apache 2.0 weights on HuggingFace. The 2B scores 83.1% on a five-benchmark panel versus Jev's published 83.0%, while deciding in ~28 ms locally. Training used only local hardware and synthetic data from a local teacher model, replicating TypeSafe's Jev experiment with small models.

A developer fine-tuned Qwen3.5 0.8B/2B and Gemma models into zero-shot classifiers ('Jeff models') that return calibrated probabilities in one forward pass, releasing Apache 2.0 weights on HuggingFace. The 2B scores 83.1% on a five-benchmark panel versus Jev's published 83.0%, while deciding in ~28 ms locally. Training used only local hardware and synthetic data from a local teacher model, replicating TypeSafe's Jev experiment with small models.

Releases·Tool Release

I open-sourced CasaJev: a Jev agent loop that builds and reuses verified tools

A developer released CasaJev, an Apache-2.0 open-source local harness for TypeSafe's Jev that lets an agent build small tools when none fit a task, then register and reuse them on later tasks. In one example, building a tool took 29.29 s versus 2.07 s to reuse it. It is an early prototype with limited OS validation and the author is seeking contributors and feedback.

A developer released CasaJev, an Apache-2.0 open-source local harness for TypeSafe's Jev that lets an agent build small tools when none fit a task, then register and reuse them on later tasks. In one example, building a tool took 29.29 s versus 2.07 s to reuse it. It is an early prototype with limited OS validation and the author is seeking contributors and feedback.

Social·Community

Benchmark: TypeSafe's Jev beats DeepSeek V4.1 Flash and grep on 470-task CodeSearchNet agent code search

A developer benchmarked TypeSafe's small System 1 model Jev (via his open-source jevpipe CLI) against grep and DeepSeek V4.1 Flash on 470 CodeSearchNet Challenge searches across five languages. Jev scored the best F1 (0.67 vs 0.64 and 0.52), with better calibration, roughly twice the speed, and slightly lower cost than DeepSeek. The result matters for coding-agent design, suggesting model-based semantic search over every function can outperform grep-based lookup.

A developer benchmarked TypeSafe's small System 1 model Jev (via his open-source jevpipe CLI) against grep and DeepSeek V4.1 Flash on 470 CodeSearchNet Challenge searches across five languages. Jev scored the best F1 (0.67 vs 0.64 and 0.52), with better calibration, roughly twice the speed, and slightly lower cost than DeepSeek. The result matters for coding-agent design, suggesting model-based semantic search over every function can outperform grep-based lookup.