开源模型LongCat-2.0在Duck Hunt游戏编程测试中与GPT-5.5表现相当，但成本为零。测试由atomic.chat使用kilocode CLI的agent完成，双方均需在三次agent迭代内构建并改进同一游戏（含鸭子、波浪、弹药、物理碰撞、掉落动画和狗回收循环）。LongCat-2.0本地运行70.3K token（$0），GPT-5.5云端运行64.9K token（$0.65）。美团将其列为1.6T参数MoE，每token约48B active。结果证明，对于小型明确任务，本地开源模型质量已可接近前沿云模型，主要差异转向成本。
So cool, Open-source model LongCat-2.0 matched GPT-5.5 on a Duck Hunt coding run for $0.
Test was done by atomic[.]chat, a desktop app that runs LLMs locally using @kilocode CLI with their agent.
The side-by-side run used 70.3K tokens locally against 64.9K cloud tokens costing $0.65.
The task was not a prompt answer; the agent had to build and revise code.
LongCat apparently handled ducks, waves, ammo, hit physics, falling animation, and the dog retrieval loop well enough to look competitive in a three-iteration agent workflow.
Meituan lists LongCat-2.0 as a 1.6T-parameter MoE with about 48B active per token.
Shows something very practical: for small, clearly defined tasks, a local open model can sometimes produce work that looks almost as good as a frontier cloud model.
So the main difference may stop being quality and start being cost.
Open-weight LongCat 2.0 matched GPT-5.5 level on agentic game dev for $0! We ran Meituan's LongCat 2.0 against cloud frontier GPT-5.5 in @kilocode CLI with their agent. Same task for both - build a retro Duck Hunt game in one game.html, improved over 3 agent iterations with duck waves, ammo and physics Outputs: LongCat 2.0: 70.3K tokens, $0.00 GPT-5.5: 64.9K tokens, $0.65 LongCat kept up on graphics, physics and game logic. Ducks fly and fall when hit, the dog fetches them, ammo counts down, the waves keep coming. Both ran clean and nothing clipped. The only difference was the bill - GPT cost $0.65, LongCat ran local for $0在 X 查看被引用的帖子
来源：Rohan Paul · x.com