【帖子标题】：GLM 5.3 and GLM 5.3 Flash ran locally on RTX PRO 6000 WS and built a penthouse using BlenderMCP 【帖子标题】：在RTX PRO 6000 WS工作站本地运行GLM 5.3和GLM 5.3 Flash，通过BlenderMCP构建顶层公寓
【帖子正文】： I keep seeing demos of AI agents building scenes in Blender through BlenderMCP, so I tried it myself. I ran both models locally for this and picked the GLM 5.3 family(Q4 quant) because videos of it doing 3D work kept showing up in my twitter feed (out of curiosity, I ran the same prompt through the full GLM 5.3, also locally with a Q4 quant) 我经常看到AI智能体通过BlenderMCP在Blender中构建场景的演示，于是决定亲自尝试。我在本地运行了这两个模型，选择GLM 5.3系列（Q4量化版本），因为我的推特时间线上不断出现它处理3D工作的视频（出于好奇，我用完整版GLM 5.3也跑了相同的指令，同样采用Q4量化本地运行）
these aren’t small models, obviously, a 4-bit quantized Flash is around 190-200GB + headroom for context. full GLM 5.3 is around 450-470GB at 4-bit quantization (basically I went with the Q4 quants for both and the RTX PRO 6000 WS GPU, though I had to rent 4x rtx pro 6000ws for the flash model and 6x for the base one) 这些模型体积不小——4bit量化的Flash版本约190-200GB，还需预留上下文空间；完整版GLM 5.3在4bit量化下约450-470GB（我同时使用了Q4量化版本和RTX PRO 6000 WS显卡，不过运行Flash模型需要租用4块显卡，基础版则需要6块）
writing the prompt wasn’t as easy as I thought. my first attempts were vague and mostly produced 3D goo instead of an actual room. I eventually started specifying real dimensions: ceiling heights, stair rise, window mullion spacing and so on(the camera work was separately done by claude opus 5 so that I wouldn’t have my token stats inflated by it) 编写指令比想象中困难。最初几次尝试过于模糊，生成的更像是3D粘液而非真实房间。后来我开始指定具体尺寸：天花板高度、楼梯踏步高、窗棂间距等（镜头运动由Claude Opus 5单独处理，避免增加token统计量）
prompt 指令内容：
model a luxury duplex penthouse in the open Blender session. footprint 20.0 x 13.0 m (260 sqm). main ceiling 2.9 m. a double-height volume 9.0 x 8.0 m rising to 6.2 m. mezzanine floor at 3.1 m with a 1.1 m balustrade. stair: 17 treads, rise 0.182, going 0.28. terrace 20.0 x 4.5 m at Z = -0.02 with a 1.15 m balustrade. curtain wall with mullions every 1.5 m, frame depth 0.06. doors 2.10 m. counters 0.90 m. dining table 0.74 m. sofa seat 0.42 m. materials, PBR ranges: glass IOR 1.45-1.52, transmission 1.0; concrete roughness 0.25-0.40; marble roughness 0.08-0.15; brushed metal metallic 1.0, roughness 0.25-0.35; fabric roughness 0.75-0.95. reference real penthouses for proportion. furnish it. do NOT add a camera. do not reset the session. 在打开的Blender会话中建模豪华复式顶层公寓。占地面积20.0x13.0米（260平方米）。主天花板2.9米。挑高空间9.0x8.0米，高度6.2米。夹层地板高度3.1米，配1.1米栏杆。楼梯：17级踏步，踏步高0.182，踏面宽0.28。露台20.0x4.5米（Z=-0.02），配1.15米栏杆。幕墙每1.5米设竖梃，框深0.06米。门高2.10米。柜台0.90米。餐桌0.74米。沙发座高0.42米。材质PBR参数：玻璃折射率1.45-1.52，透射率1.0；混凝土粗糙度0.25-0.40；大理石粗糙度0.08-0.15；拉丝金属金属度1.0，粗糙度0.25-0.35；织物粗糙度0.75-0.95。参考真实顶层公寓比例。布置家具。不要添加摄像机。不要重置会话。
at first it was putting up the curtain wall, stairs, mezzanine, the glass railing, all that, then at some point I noticed it had furnished the place too with some furniture: sofa, dining table and plates on it. the pendant lights were hanging from these 4 m cords, and for some reason it had modeled the individual spines on the books, which I never asked for 最初它搭建了幕墙、楼梯、夹层、玻璃栏杆等结构，后来我注意到还添加了家具：沙发、餐桌及餐具。吊灯垂挂着4米长的电线，不知为何连书本的书脊细节都建模了——这并非我的要求
the video only follows the camera through the living space, so the terrace and facade aren’t visible(the clip is repurposed from another video I made with the same scene, I didn’t render a new one because that takes quite some time) 视频仅展示生活空间的镜头移动，因此露台和外立面不可见（这段素材取自同一场景的其他视频，由于渲染耗时较长，我没有重新制作）
stats 统计数据：
metric Flash GLM 5.3 指标 Flash版 GLM 5.3完整版
objects 811 847 物体数量 811个 847个
turns 43 42 交互轮次 43次 42次
tool errors 9 8 工具错误 9次 8次
thinking before 1st object 10s 21m 55s 生成首个物体前思考时间 10秒 21分55秒
time 38m 52s 40m 43s 总耗时 38分52秒 40分43秒
output tokens 36K 112K 输出token数 3.6万 11.2万
GLM 5.3 spent 22 minutes thinking(82k tokens), before placing any objects(as well as producing 36 more objects than GLM 5.3 Flash and consuming 3x times the output tokens), meanwhile GLM 5.3 Flash got to work almost immediately 完整版GLM 5.3放置首个物体前思考了22分钟（消耗8.2万token），最终比Flash版多生成36个物体且消耗3倍输出token，而Flash版几乎立即开始工作
I measured both scenes afterwards by raycasting upward from the floor and checking the rooms against the brie. Flash got the double-height void right at 9 x 8 m. the full model built it at 9 x 4.5 m but reported it as 9 x 8 m 事后我通过地面向上射线检测验证空间尺寸：Flash版准确实现了9x8米的挑高空间，而完整版实际建成9x4.5米却报告为9x8米
This is obviously just an experiment, not a benchmark. Flash came surprisingly close on object count and total time while using less than one-third as many output tokens. it also got the main room dimensions right when the full model didn’t 这显然只是实验而非基准测试。Flash版在物体数量和总耗时上表现接近，却只用了不到1/3的输出token，而且在主房间尺寸上比完整版更准确
if you want to try the same Blender setup, I used 若想尝试相同Blender配置，我使用了：
the community BlenderMCP project 社区版BlenderMCP项目
I’m a founder of atomic.chat, we have an app for running local models and our own quants(any feedback is appreciated, we’re trying to make our products as good as possible for you guys) 我是atomic.chat的创始人，我们开发了运行本地模型的应用程序并提供自主量化版本（欢迎反馈，我们致力于为用户打造最佳产品）
🔗 知识库双向关联
- [[Raw_翻译_[AINews] Memory prices up 500 in 12 months|Memory prices up 500 in 12 months_全翻译]]
- [[Auto_简报_[AINews] Death of Params Z.ai CEO Jie Tang on GLM 5.3 and th|Death of Params Z.ai CEO Jie Tang on GLM 5.3 and th_简报]]
- Using Local Coding Agents_全翻译