菜单
Download GGUF models from HuggingFace directly into LM Studio's model directory. Triggers on "download model", "get model", "grab model", "install model".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Orchestrate LiteBench model testing — scan LM Studio for models, run the agent harness against each, collect scores, swap models, and produce a leaderboard. Triggers on "test models", "run benchmark", "run harness", "test all models", "benchmark models", "start testing".
Autonomous agent and skill trainer — transforms the current Claude session into a self-improving loop inspired by Karpathy's autoresearch. Activates on: "train", "evolve agent", "improve agent", "run training loop", "agent trainer", "evolve this agent", "train this agent", "self-improving agent", "auto-improve", "training loop", "agent evolution", "autotrainer", "improve skill", "evolve skill", "optimize skill", "skill training", "improve my skills". Targets agents (.claude/agents/) OR skills (.claude/skills/) — auto-detects which. For agents: spawns A/B variants, evaluates output quality, mutates config. For skills: uses Claude Code's built-in eval system (run_eval.py) to measure trigger accuracy, then mutates the skill description. Use when you want an agent or skill to improve itself over N cycles without manual intervention.
Tune the agent harness system prompts to improve model tool-calling scores. Triggers on "tune harness", "improve scores", "train harness", "optimize prompts".
| name | model-download |
| description | Download GGUF models from HuggingFace directly into LM Studio's model directory. Triggers on "download model", "get model", "grab model", "install model". |
| type | skill |
Download GGUF model files from HuggingFace into LM Studio's model directory so they can be loaded and tested with LiteBench.
Models go to: ~/.lmstudio/models/<publisher>/<model-name>/
On Windows: C:/Users/<username>/.lmstudio/models/
curl -s "https://huggingface.co/api/models?search=<query>+gguf&limit=10" | python -c "import sys,json; d=json.load(sys.stdin); [print(m['id']) for m in d]"
curl -s "https://huggingface.co/api/models/<repo-id>/tree/main" | python -c "
import sys, json
files = json.load(sys.stdin)
for f in files:
if f.get('path','').endswith('.gguf'):
size_mb = f.get('size', 0) / 1024 / 1024
print(f'{f[\"path\"]:60s} {size_mb:.0f} MB')
"
Prefer Q4_K_M quantization (good balance of quality and size).
mkdir -p ~/.lmstudio/models/<publisher>/<model-GGUF>/
curl -L -o ~/.lmstudio/models/<publisher>/<model-GGUF>/<filename>.gguf \
"https://huggingface.co/<repo-id>/resolve/main/<filename>.gguf"
ls -lh ~/.lmstudio/models/<publisher>/<model-GGUF>/
The model will appear in LM Studio after a rescan. User needs to load it manually in LM Studio's UI.
For tool calling / agent benchmarking:
| Model | Repo | Size | Why |
|---|---|---|---|
| xLAM 2 1B | Salesforce/xLAM-2-1b-fc-r-gguf | ~940MB | Purpose-built for function calling |
| xLAM 2 8B | Salesforce/xLAM-2-8b-fc-r-gguf | ~5GB | #1 on BFCL leaderboard |
| Hermes 3 8B | NousResearch/Hermes-3-Llama-3.1-8B-GGUF | ~5GB | Gold standard structured tool use |
| Llama 3.1 8B | lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF | ~5GB | Meta native tool calling |
| Llama 3.2 3B | lmstudio-community/Llama-3.2-3B-Instruct-GGUF | ~2GB | Smallest Llama with tools |