---
{"schema":1,"date":"2026-09-17","timezone":"Asia/Shanghai","title":"AI 原生游戏引擎每日观察","revision":1,"publication":"publish","review_status":"generated","summary":["Zing补齐论文和实时服务，键盘与文字可共同控制。","小型神经运行时在CPU上可玩，但低清且很快漂移。","CoreAI接通Unity多人链路，世界状态同步仍未解决。"],"events":[{"project":"zing-0-5","title":"Zing-0.5发布技术报告并公开模型与推理代码","kind":"paper","event_date":"2026-09-15","publication_date":"2026-09-15","discovered_date":"2026-09-17","evidence":"source-claim","source":"https://arxiv.org/abs/2609.17909"},{"project":"zing-0-5","title":"Zing-SGLang加入实时推理与WebSocket服务","kind":"code","event_date":"2026-09-14","publication_date":null,"discovered_date":"2026-09-17","evidence":"verified","source":"https://github.com/seedleap/Zing-SGLang/commit/60ee3302606efcbd04980bfe5140c841c57d51ad"},{"project":"neural-game-engine","title":"神经游戏引擎完成2M至26M扩展实验和CPU基准","kind":"code","event_date":"2026-09-15","publication_date":null,"discovered_date":"2026-09-17","evidence":"source-claim","source":"https://github.com/GaganGutta/neural-game-engine/commit/4c3692aed3e193adf5e126958495a56f6e73de62"},{"project":"coreai","title":"CoreAI 7.42把Mirror多人传输接入Unity场景","kind":"code","event_date":"2026-09-16","publication_date":null,"discovered_date":"2026-09-17","evidence":"verified","source":"https://github.com/NeoXider/CoreAI/commit/85dce51de80b0f0b79d2297a8715441479f25886"},{"project":"coreai","title":"CoreAI 7.44.1修复场景销毁时的取消异常","kind":"code","event_date":"2026-09-17","publication_date":null,"discovered_date":"2026-09-17","evidence":"verified","source":"https://github.com/NeoXider/CoreAI/commit/de077ddeff07c14276e4a08cefbfe4e964d3a72b"}]}
---
# AI 引擎日报｜2026-09-17

**今天一句话：** Zing把实时世界模型从演示推进到可部署服务，但状态、物理和长期一致性仍未脱离传统引擎。

## 今天值得看

### Zing-0.5：论文与实时服务补齐

**新在哪：** 9月14日上线SGLang实时服务，15日公开技术报告；它能在同一段生成中同时接收键盘动作和文字改场景。**有什么用：** 可快速试做能操作、能边玩边改的画面流。**现在能用吗：** 模型、推理和WebSocket代码都已开放；作者称特定服务器条件下可跑832×480、24帧/秒，约0.009美元/分钟，尚无独立复测。长时间运行会漂移，也没有游戏逻辑和可靠物理。

[论文 · 2026-09-15](https://arxiv.org/abs/2609.17909) · [实时服务代码 · 2026-09-14](https://github.com/seedleap/Zing-SGLang/commit/60ee3302606efcbd04980bfe5140c841c57d51ad) · [模型权重 · 2026-08-26](https://huggingface.co/seedleap/zing-0.5)

### [开源项目] Neural Game Engine：小模型在CPU上跑到可操作

**新在哪：** 9月15日完成从200万到2570万参数的扩展实验，并放出可直接玩的200万参数检查点。**有什么用：** 开发者能在普通电脑验证“只靠动作预测下一帧”的低成本路线。**现在能用吗：** 作者在自己的笔记本CPU上测得小模型35.47帧/秒；但画面只有64×64，需先喂6帧真实《Doom》画面，约十几帧后就会漂移，较大模型权重也未提供。

[代码与检查点 · 2026-09-15](https://github.com/GaganGutta/neural-game-engine) · [CPU基准提交 · 2026-09-14](https://github.com/GaganGutta/neural-game-engine/commit/ddf9dd3e31f403d0d64d4597b6faabc0276fe8f1)

### CoreAI：Unity多人Agent接线可从场景启用

**新在哪：** 9月16日的7.42源码提交让Mirror多人传输首次可由Unity场景直接启用，17日又修复界面销毁时的取消异常。**有什么用：** 做多人Agent原型时少写一层手工接线。**现在能用吗：** 源码可取，但没有对应的新GitHub Release；世界状态仍不走网络同步，且采用非商业许可，不能当作成熟商用联网方案。

[多人传输提交 · 2026-09-16](https://github.com/NeoXider/CoreAI/commit/85dce51de80b0f0b79d2297a8715441479f25886) · [修复提交 · 2026-09-17](https://github.com/NeoXider/CoreAI/commit/de077ddeff07c14276e4a08cefbfe4e964d3a72b)

## 我的看法

今天三条都指向同一缺口：能生成还不够，真正难的是确定状态、长期一致性和联网同步。

观察窗口：2026-09-10至2026-09-17。官网/仓库、国内/中文与论文已查；X仅检索公开索引，未找到可回溯的及时线索。