Fireworks 让 Qwen 跑 11 小时智能体循环
Fireworks AI 为 Qwen 模型新增长时智能体循环功能，支持观察、推理、编码、执行和验证的重复流程。Qwen 官方演示运行了 11 小时，生成了超过 10,000 行代码并执行了 1,000 多次调用。Fireworks 提供 reasoning_history 参数以跨轮次保留推理上下文，支持按请求切换思考/非思考模式，以及原生图像和文本输入。默认启用 262k 上下文和提示缓存，缓存输入价格为每百万 token 0.10 美元。
Built for long-horizon agent loops. Observe → reason → code → act (GUI/CLI) → verify, repeat. Qwe...
Built for long-horizon agent loops. Observe → reason → code → act (GUI/CLI) → verify, repeat. Qwen’s own demo ran 11 hours with 10k+ lines of code and 1k+ calls. On Fireworks you get: → reasoning_history="preserved" to persist reasoning context across turns → Per-request thinking / non-thinking mode → Native image + text input → 262k context + prompt caching on by default ($0.10/1M cached input) 💬 1 🔄 0 ❤️ 0 👀 853 📊 1 ⚡