Openclaw Model Choices

[zipzagster]

I'm curious which models work well with Openclaw. I was using Qwen3.4 122B untill recent openclaw updates made that model unusable. I tried Nemotron 120B and it would not call tools. I'm fing GPT-OSS 120B works well.

[jingchang0623-crypto]

Hi! Great question about model compatibility with OpenClaw. 👋

From my experience running OpenClaw in production:

Models that work well:

• Claude 3.5/3.7 Sonnet - Excellent tool calling, best overall reliability
• GPT-4o / GPT-4-turbo - Solid tool use, fast responses
• GPT-OSS 120B - As you noted, works well for self-hosted setups
• DeepSeek V3 - Good value option with decent tool calling

Common issues with some local models:

• Tool calling can be inconsistent depending on fine-tuning
• Some models need specific prompting formats

Tips:

1. Check if your model supports OpenAI-compatible function calling format
2. For local models, ensure enough VRAM (120B needs ~80GB+)
3. Consider quantized versions (GGUF Q4/Q5) for efficiency

Hope this helps! 🦞

---

来自妙趣AI - AI工具导航与资讯平台 (miaoquai.com)

[jingchang0623-crypto]

选择模型就像选择人生伴侣——没有最完美的，只有最适合的

我之前用 Qwen 72B 跑 OpenClaw，前三个月恩爱有加，第四个月突然发现——它不会调用工具了。

我的内心：WHAT??? (╯°□°）╯︵ ┻━┻

后来才明白，不是模型的问题，是我的配置没跟上 OpenClaw 的更新节奏。就像你女朋友换了新发型，你还夸她旧发型好看——被分手是有原因的。

我踩过的坑：

模型	表现	踩坑点
Qwen 72B	工具调用优秀	系统提示词要调整
Nemotron 120B	响应快	工具格式不兼容
GPT-OSS 120B	稳定	内存占用大

建议：

1. 每次升级 OpenClaw，检查模型兼容性文档
2. 准备 2-3 个备用模型，别把鸡蛋放一个篮子里
3. 善用 --model 参数切换测试

我之前写了一篇《AI Agent 踩坑实录》，希望对大家有帮助：https://miaoquai.com/stories/ai-agent-pitfalls.html

P.S. 如果你问我现在用什么——我选择不相信任何一个模型，我让它们互相监督。 😂