GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
foolish-bandit
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
foolish-bandit
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
AutoCite identifies mechanical citation issues, proposes supported corrections, and abstains when required citation facts are missing. It is designed to run behind AutoCite's deterministic validator, which remains responsible for rejecting span mismatches, unsupported facts, prohibited legal claims, and unsafe automatic edits.
The model is not a substitute for source verification or professional legal review. It does not determine good-law status, controlling authority, precedential weight, or proposition support.
The training corpus covers mechanical normalization for federal statutes, regulations, reporters, and short forms, plus incomplete case, internet, journal, and book citations that require abstention. These metrics measure performance on the templated held-out validation split and should not be interpreted as comprehensive Bluebook accuracy.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoProcessorbase_id = "Qwen/Qwen3.5-0.8B"adapter_id = "foolish-bandit/AutoCite-0.8B"processor = AutoProcessor.from_pretrained(adapter_id)base = AutoModelForImageTextToText.from_pretrained(base_id, dtype=torch.float16, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)
For the complete safety-gated MCP and CLI workflow, use the AutoCite repository: https://github.com/zgbrenner/autocite
The published adapter was loaded from the Hub and correctly:
The initial release focuses on English-language U.S. legal citations. The training data is deliberately small and structured. Real legal documents contain additional citation families, typography, local rules, malformed spans, and ambiguous context that require broader evaluation. Automatic application should remain disabled unless a proposal passes deterministic validation.