GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ZaoKing
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ZaoKing
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Do not use Khudi for medical, legal, financial, emergency, identity-verification, employment, or other high-stakes decisions. It can be wrong, incomplete, biased, or unsafe. In an emergency, contact local emergency services or a qualified professional rather than relying on a model response.
This is not a claim that Khudi reliably covers every Pakistani language, dialect, domain, or fact. Pashto, Sindhi, Saraiki, Balochi, legal, and finance behaviour still need native-speaker or domain-expert review.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfigbase_id = "Qwen/Qwen3.5-9B"adapter_id = "ZaoKing/khudi-v1-0-release-gate-repair-candidate"processor = AutoProcessor.from_pretrained(base_id)model = AutoModelForImageTextToText.from_pretrained(base_id,device_map="auto",dtype=torch.bfloat16,quantization_config=BitsAndBytesConfig(load_in_4bit=True),)model = PeftModel.from_pretrained(model, adapter_id)model.eval()
For direct answers without visible reasoning, use the Qwen chat template with enable_thinking=False where supported by your installed Transformers version.
Khudi V1.0 is a continuation of the broader Khudi V2 fine-tune, whose preserved core training corpus contains 42,406 instruction-response records (with a separate 184-record extension). This V1.0 release then applies a focused repair fine-tune on 202 accepted instruction-response records: 178 approved rehearsal records and 24 targeted release-gate repairs.
The final 202-record repair set contains English, Urdu-script, and Roman Urdu examples, with emphasis on:
The adapter was trained with 4-bit QLoRA on an RTX 4090. This focused repair run is not a substitute for broad pretraining, expert validation, or a public benchmark.
The current adapter passed a six-prompt targeted regression check covering the repaired categories above. That is a smoke check, not a broad capability score.
The full 25-case internal PakBench seed evaluation was completed for this exact adapter using deterministic direct-answer generation (max_new_tokens=256). This remains a provisional internal assessment, not a public benchmark score:
The completed seed does not establish broad quality, production readiness, or reliable finance/legal/medical capability. PakBench is currently an internal research seed, not a finished public benchmark. See eval/runs/khudi-v1-0-release-full-pakbench/PROVISIONAL_ASSESSMENT.md in the project source for the item-level internal assessment.
Please report failures, especially Urdu and Roman Urdu failures, with:
Use the repository Discussions tab for feedback. Do not post CNICs, phone numbers, private documents, credentials, or other personal information.
Khudi is built on Qwen/Qwen3.5-9B. Use of the base model and this adapter remains subject to the applicable Qwen terms and any downstream deployment requirements. No separate Khudi license is asserted in this research-preview repository.
bibtex
@software{khudi_v1_0_research_preview,title = {Khudi V1.0 Research Preview},author = {Khudi AI},year = {2026},url = {https://huggingface.co/ZaoKing/khudi-v1-0-release-gate-repair-candidate}}