GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
vXofi
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
vXofi
Model Tree
huihui-ai/Huihui-Qwen3.5-9B-abliterated
Input Modalities
Output Modalities
Supported Functionality
The model explores small-data domain adaptation for short, slang-heavy Telegram-style conversation. It is intended for research and private, allowlisted informal-chat deployments with human oversight.
It is not intended as a general assistant, factual authority, moderation model, or safety-critical system.
5e-5;The raw training corpus and generated augmentation are private and are not included in this repository.
| Comparison | Result | 95% session-cluster CI |
|---|---|---|
| v16 artifact vs unadapted base artifact, 40 contexts | v16 preference 80.0% | 67.9%-88.9% |
| v15 vs v16, 64 contexts | v15 preference 61.7% | 49.3%-73.3% |
| production vs legacy prompt, 43 contexts | production preference 55.8% | 45.0%-65.3% |
On a separate 43-context absolute audit, one primary reviewer judged 51.2% of
responses good and another 37.2% acceptable: 88.4% were therefore
reviewer-usable. This is an ordinal single-reviewer result, not a production
acceptance rate.
The base comparison is artifact-level rather than a template-identical weight-only ablation: base and v16 use their intended native no-thinking templates. v15 is directionally preferred to v16, so v16 is not presented as an across-the-board quality upgrade.
Full methodology, definitions, confidence intervals, and limitations: public evaluation report.
Raw conversations, evaluation contexts, model responses, and human ratings are not public. Public reports contain aggregate statistics only. The current public experiment manifest represents operational prompts with neutral descriptions and SHA-256 hashes; prompt secrecy is not a security boundary.
Do not expose this model through an unrestricted public endpoint. Add authentication, allowlisting, rate limits, and content controls appropriate to the deployment.