GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Mhaquehaque
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
Mhaquehaque
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
| Method | QLoRA: NF4 base, fp16 compute, fp32 LoRA, r=32 / alpha 64 on all linear layers, DDP on 2× Kaggle T4 |
| Data | berkelium-cad-v1: 18,712 core-verified rows (33 part families + feature grammar + gear/beam CEMs + lantern); random 12,000-row training subset, 15.5M tokens |
| Steps | 601 (0.80 epoch, stopped by a 10 h time guard); 10.06 h |
| Loss | train 1.03 → 0.005; val 0.0042 (120 in-distribution rows) |
| Data manifest | sha256 bb07a39a…fe49d |
| Adapter | adapter_model.safetensors sha256 c76b64ee…2f12f (174.7 MB, fp16) |
Held-out families never trained on: pipe_elbow, chair, knob, hinge_leaf.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218")model = PeftModel.from_pretrained(base, "Mhaquehaque/berkelium-qwen3-8b-cad-v1")Use the system prompt from berkelium_neural/cad/prompts.py (prompt version cad-0.1), with thinking mode off.
Then run the JSON through berkelium.pipeline.run(...). A GGUF (Q4_K_M) for Ollama/llama.cpp is published at
Mhaquehaque/berkelium-qwen3-8b-cad-v1-GGUF.
Code: github.com/enigmahacker/Berkelium-Neural-Rendering-Studio (branch Berkelium-Neural-Rendering-Studio-neural-v1).