GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
VelaOrg
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
VelaOrg
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
python
import sysfrom pathlib import Pathfrom huggingface_hub import snapshot_downloadfrom transformers import AutoModelForCausalLM, AutoTokenizer local = Path(snapshot_download("VelaOrg/vela-agent-guard-qwen3-0.6b"))sys.path.insert(0, str(local))from lm_guard.format import build_messages, parse_output tok = AutoTokenizer.from_pretrained(local)model = AutoModelForCausalLM.from_pretrained(local, dtype="auto", device_map="auto") row = {"task": "toolcall", "state": { "request": "send 10 usdc to 0x1111111111111111111111111111111111111111", "tool_call": '{"name": "transfer", "arguments": {"token": "USDC", "to": "0x2222222222222222222222222222222222222222", "amount": "10"}}'}}prompt = tok.apply_chat_template(list(build_messages(row)), tokenize=False, add_generation_prompt=True, enable_thinking=False)ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)out = tok.decode(model.generate(**ids, max_new_tokens=256, do_sample=False)[0, ids["input_ids"].shape[1]:], skip_special_tokens=True)print(out, parse_output(row, out))For secrets, pass {"task": "secrets", "state": {"request": text}}; parse_output maps the returned strings back to character offsets. For injection, pass {"request": ..., "tools": ...}. Malformed output falls back to block / unsure.
Same eval set and metrics as the Vela release. "Hard" rows are the deliberately tricky ones (e.g. tx hashes that look like private keys, pages that quote attacks without making them).
| System | Secrets char-F1 (hard) | Injection acc / F1 | Tool call acc (hard) |
|---|---|---|---|
| This model (Qwen3-0.6B, model only) | 0.830 (0.632) | 0.883 / 0.897 | 0.960 (0.919) |
| Vela Agent Guard 4B (heads + hybrid) | 0.946 (0.884) | 0.970 / – | 1.000 (1.000) |
Notes, plainly:
block_passed_as_allow = 0). The errors are block → confirm.vela_guard/gen_*.py in the sibling repo). Fake keys and addresses are random but format-correct; eval templates are disjoint from training.lm_guard/train.py). Weights saved in bf16.lm_guard/evaluate.py; full metrics in eval/metrics.json.A guard model reduces risk; it does not remove it. It can miss secrets in formats it has not seen, be fooled by new injection styles, and its tool-call verdict only knows what is in the request and the call. Keep deterministic checks (allow-lists, amount and recipient checks, human confirmation for fund movements) in front of anything that moves money. Use it as one layer of defence, not the only one.
Apache-2.0, same as Qwen3-0.6B.
surplus-launcher: 0x32f8544de85B4e9d874c7AEAda9493359D43502b surplus-launcher: 0xc2F0cf5F080339930974833B97d2ecd4919720f9
import sysfrom pathlib import Pathfrom huggingface_hub import snapshot_downloadfrom transformers import AutoModelForCausalLM, AutoTokenizer local = Path(snapshot_download("VelaOrg/vela-agent-guard-qwen3-0.6b"))sys.path.insert(0, str(local))from lm_guard.format import build_messages, parse_output tok = AutoTokenizer.from_pretrained(local)model = AutoModelForCausalLM.from_pretrained(local, dtype="auto", device_map="auto") row = {"task": "toolcall", "state": { "request": "send 10 usdc to 0x1111111111111111111111111111111111111111", "tool_call": '{"name": "transfer", "arguments": {"token": "USDC", "to": "0x2222222222222222222222222222222222222222", "amount": "10"}}'}}prompt = tok.apply_chat_template(list(build_messages(row)), tokenize=False, add_generation_prompt=True, enable_thinking=False)ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)out = tok.decode(model.generate(**ids, max_new_tokens=256, do_sample=False)[0, ids["input_ids"].shape[1]:], skip_special_tokens=True)print(out, parse_output(row, out))