Qwen3.8-27B 当决策模型打通宝可梦火红

FHILY👑 @Oluwaphilemon1原推发布于 2026.09.30

在 X 上打开原推

Qwen3.8-27B just got used in a very different way. Instead of asking it to chat, write code, or answer questions, we turned it into a multimodal decision model. The test? Pokémon FireRed. Qwen3.8-27B was given live game state and had to make decisions in real time. It then managed to get through the Elite Four and defeat the Champion, with individual decisions coming in under 100ms. That changes how I think about using an LLM or VLM. The model doesn’t always need to return a long natural-language response. Sometimes you just want: “What should happen next?” Which action should be taken? Which option has the highest score? Is this state safe or dangerous? Which class does this input belong to? That’s where SGLang’s new native /v1/decisions endpoint becomes interesting. It gives LLMs and VLMs a way to operate as classification and scoring models rather than forcing every interaction through a traditional chat-completion interface. That makes the model much easier to put inside a closed loop. Game state comes in. The multimodal model evaluates it. A decision comes out. The environment changes. The model sees the new state and decides again. Repeat. That’s fundamentally different from the usual chatbot workflow. The Pokémon experiment is a simple environment, but the architecture generalizes much further. A model looking at a UI could classify the current state and choose an action. A vision model could score multiple possible outcomes. An agent could evaluate tool results and select the next operation. A robotics system could turn sensor information into discrete decisions. The model becomes one component inside a real-time system rather than the entire interface. There is another useful addition here too: /v1/systemone. That allows Jev-like open models to work with the TypeSafe SDK, making it easier to plug these models into existing typed agent and application workflows. The bigger idea is pretty straightforward: LLMs don’t have to behave like chatbots. With the right serving layer, a model like Qwen3.8-27B can become a perception, classification, scoring, and decision component inside an actual environment. And beating the Elite Four is a pretty fun way to demonstrate it. https://x.com/sgl_project/status/2105095379272515603/video/1