Irving

捕捉每一丝灵感，将天马行空落地为产品
AI Agent Builder |
Bitcoin Core Dev |

[Irving @BlockInsight214]

You can now run GLM-5.2 locally on Mac Studio to integrate with Hermes.👇

Hardware Reality (Non-Negotiable)

• 2-bit dynamic quant: ~239GB

• Minimum: 256GB Unified Memory (Mac Studio only, no laptops)

• Recommended: 512GB Unified Memory

• Speed: 1–9 tokens/sec on M3 Ultra

Use case: Perfect private background worker for long async tasks
Not for fast, casual interactive chat

Step 1: Run GLM-5.2 Local API (2 Options)

Option 1 | LM Studio (Easiest macOS Setup)

1. Install LM Studio

2. Download Unsloth GLM-5.2 GGUF (UD-IQ2_M)

3. Enable developer local server

4. Endpoint: http://localhost:1234/v1

Option 2 | Llama.cpp (Full CLI Control)
pip install huggingface_hub

hf download unsloth/GLM-5.2-GGUF \
--local-dir unsloth/GLM-5.2-GGUF \
--include "*UD-IQ2_M*"

./llama.cpp/llama-server \
--model unsloth/GLM-5.2-GGUF/UD-IQ2_M/GLM-5.2-UD-IQ2_M-00001-of-00006.gguf \
--temp 1.0 --top-p 0.95 --min-p 0.01 \
--ctx-size 32768 --jinja \
--host 0.0.0.0 --port 8080
(Official Unsloth sampling params + Jinja chat template for valid tool calling)

Step 2: Connect Nous Hermes Agent to Local Model

Edit ~/.hermes/config.yaml for fully local agent execution:
model:
default: glm-5.2
provider: custom
base_url: http://localhost:8080/v1
api_key: local
context_length: 32768
agent:
tool_use_enforcement: true

Key fix: Enable tool_use_enforcement
GLM is not in Hermes’ default supported model list — this forces proper tool calling (no more just describing tasks!)

[Irving @BlockInsight214]

@brian_armstrong The vision is clear, but the real difficulty is building an agent that can handle real money movement, unexpected market moves, and regulatory edge cases without creating silent disasters that only surface days later.

[Irving @BlockInsight214]

I've been running similar multi-agent setups on bigger refactors and the skeptic + reviewer layer is what actually keeps things from quietly drifting into elegant but broken solutions — the real test will be how well /goal handles mid-project plan changes without the whole team losing coherence.

[Irving @BlockInsight214]

@yuhasbeentaken The token inefficiency and long planning loops are the real hidden cost — even with much lower per-token pricing, GLM-5.2 can easily end up more expensive than expected once you run actual long-horizon agent workflows that need consistent steering.

[Irving @BlockInsight214]

I'm desperately in need of GLM-5.2 right now.

My current agent workflows are burning through GPT-5.5 tokens way too fast, and the monthly bill is getting out of control.

[Irving @BlockInsight214]

@Hicker_Moledao This is the exact kind of infrastructure oversight that hurts long-running local agents the most — TRACE-level logging should never be left on by default in anything meant to run for hours, or you end up trading model intelligence for hardware lifespan.

[Irving @BlockInsight214]

@0xgibly I've wasted too many tokens and context windows on agent hallucinations that actually traced back to messy PDF parsing — adding a proper cleaning step upstream has been one of the highest-ROI improvements in any document-heavy workflow I've run.

[Irving @BlockInsight214]

🗂️ 论文、合同、扫描件丢给 AI 之前，最难的一步往往是「先把 PDF 洗干净」。这几个开源项目专干这件事：转成 Markdown/JSON，直接喂给 RAG 或 agent。

① 📄 MarkItDown · 微软出品，Office/PDF/图片一键转 Markdown，⭐ 15万+，格式覆盖最全
🔗 https://t.co/4jxUro8cZ6

② 🧪 MinerU · 复杂 PDF/Office 转 LLM 友好 Markdown/JSON，⭐ 6.8万，论文和研报解析很能打
🔗 https://t.co/xBbOkMm3dN

③ 🤖 Docling · IBM 开源，文档清洗后接 gen AI pipeline，⭐ 6.1万，企业文档场景友好
🔗 https://t.co/aem1SAv5hM

④ ✂️ marker · 高精度 PDF → Markdown + JSON，⭐ 3.6万、GPL-3.0，扫描版也能救
🔗 https://t.co/36vId9A7mu

⑤ 🔍 surya · OCR + 版面分析 + 表格识别，90+ 语言，⭐ 2万、Apache-2.0，复杂排版先过它
🔗 https://t.co/7wvhvZYuU9

💡 怎么选
👉 日常 Office/PDF 快速转文本 → MarkItDown
👉 论文/研报/复杂版式 → MinerU 或 marker
👉 接企业 RAG 流水线 → Docling
👉 扫描件、表格多、排版乱 → surya 预处理 + MinerU

[Irving @BlockInsight214]

@_RyanFree 请回

[Irving @BlockInsight214]

停了两天会员，

跑了这么多粉丝，

蓝V真是试金石

[Irving @BlockInsight214]

Real agentic benchmarks like GDPval-AA matter more than most synthetic ones because they test multi-turn practical deliverables. GLM-5.2 reaching #3 here means open weights are now close enough that long-running agent workflows can start shifting for cost and control instead of pure capability.

[Irving @BlockInsight214]

@lnyisn 确实就你一个人知道

[Irving @BlockInsight214]

@YIYONG19680604 我得搞宕机让大家睡个好觉

[Irving @BlockInsight214]

由于我今天过于活跃

X刚才宕机了

[Irving @BlockInsight214]

@XChatScout 我之前在真实中大型代码库上跑长任务时，最头疼的就是后半段上下文失忆和工具调用链断掉，GLM-5.2 在 1M 窗口下还能保持相对稳定的架构分析和模块重构建议，对需要长时间自主执行的 agent 工作流来说确实是目前开源里比较能打的选项。

[Irving @BlockInsight214]

@ool69loo 中文圈喜欢短平快，

感觉继承了国人消费短视频的特点。

[Irving @BlockInsight214]

@youngstockuser 我希望他飞又怕他飞

[Irving @BlockInsight214]

@FirstFour01 这些操作其实比蓝V互关更难坚持，
尤其是“别人评论了你的帖子一定要回访”这一条

[Irving @BlockInsight214]

@cdexsta 哇，还不如闲鱼靠谱吗

[Irving @BlockInsight214]

@lixinbao_X 难道黄推不是机器人