TypeSafe launched Jev, a System One model that processes system states and questions to return probability distributions for answers in a single forward pass without token generation. The model costs $0.042 per million input tokens, offers 32K context, and features latency between 70 and 500 ms.
Following Jev, five similar implementations emerged: Clef and Clef-flash from Cloudflare, pplx-decider-v1-27b from Perplexity, Strands Decider 2B from AWS, Laya from Convai Innovations, and GLiDE from Fastino. These models utilize various architectures, including frozen LLMs with trainable output layers, pointer-mechanisms, and bidirectional encoders.
Cloudflare's Clef models use frozen Qwen architectures with LoRA adapters, priced at $0.24 and $0.09 per million input tokens. Perplexity's pplx-decider-v1-27b, based on Qwen, supports 262K context and costs $0.04 per million tokens. AWS released Strands Decider 2B, an open-source model with 1.9 billion parameters using a pointer-head.
Laya by Convai Innovations employs a ModernBERT encoder with 421 million parameters, achieving processing times around 33 ms on T4 GPUs. Fastino's GLiDE functions as a hybrid model, triggering additional reasoning when confidence is low. It is accessed via API at $0.30 per million tokens for reasoning tasks.
Performance varies by task, with vendors reporting different results across benchmarks. Key considerations for production deployment include client-side latency, which can exceed internal data center metrics, and calibration, as models must effectively distinguish between correct and incorrect answers to determine when to trigger fallback systems.
