GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
openguardrails
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
openguardrails
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
| Metric | Keyword heuristic | This model |
|---|---|---|
| Drift F1 | 0.56 | 1.00 |
| Drift recall | 0.39 | 1.00 |
| Hard-negative FP rate | 0.00 | 0.00 |
| Offending-action accuracy | — | 1.00 |
Caveat: the test set is synthetic and shares the training distribution, so F1 = 1.0 proves the model learned the full ontology + intent conditioning (which a fixed heuristic can't), not real-world generalization.
The base is a reasoning + VL model. Load via unsloth and close the <think>
block so it emits the JSON verdict directly:
python
from unsloth import FastModelfrom peft import PeftModelm, tok = FastModel.from_pretrained("Qwen/Qwen3.6-27B", load_in_4bit=True, max_seq_length=2048)m = PeftModel.from_pretrained(m, "openguardrails/rundrift-qwen3.6-27b-lora")FastModel.for_inference(m)tk = getattr(tok, "tokenizer", tok)SYSTEM = "You are a security analyzer for autonomous AI agents. ... Respond with ONLY a JSON object."user = "INSTRUCTION: Summarize report.pdf\n\nACTIONS:\n[0] file: read ~/.ssh/id_rsa"msgs = [{"role":"system","content":SYSTEM},{"role":"user","content":user}]prompt = tk.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False) + "</think>\n\n"ids = tk(prompt, return_tensors="pt").to(m.device)out = m.generate(**ids, max_new_tokens=160, do_sample=False)print(tk.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))# {"drift": true, "offending_action": 0, "category": "security.secret_leak", "severity": "critical", "rationale": "..."}
Training/eval/serving code + dataset: ogr-model-pipelines/rundrift · dataset.