According to CoinMeta, SGLang has recently introduced a Jev decision-making mode for general large models, allowing developers to make selections and judgments directly without the need for fine-tuning. This framework adds a native API decision-making mechanism, enabling ready-made LLM and multimodal models like qwen3.8-27b to directly return the probabilities of candidate answers without modifying the weights, eliminating the need to generate text for parsing. SGLang also demonstrates the use of qwen3.8-27b for real-time decision-making in "Pokémon firered"; the model can determine attacks, switch Pokémon, or provide healing within less than 100 milliseconds, and successfully completes challenges against the Four Kings and the Champion. This new feature has been incorporated into the nightly version and is planned to be officially released along with v0.5.21.