How to use from
Unsloth StudioInstall Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for squaredcuber/forge-optimizer-qwen3.6-35b-a3b to start chattingUsing HuggingFace Spaces for Unsloth
# No setup required# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for squaredcuber/forge-optimizer-qwen3.6-35b-a3b to start chattingLoad model with FastModel
pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
model_name="squaredcuber/forge-optimizer-qwen3.6-35b-a3b",
max_seq_length=2048,
)Quick Links
forge-optimizer
A LoRA fine-tune of Qwen3.6-35B-A3B (MoE) that turns unoptimized backend code into efficient, correct code — the skill measured by forger-bench, an efficiency-aware benchmark for AI-generated InsForge SDK code.
Trained with Unsloth (bf16 LoRA) + an agentic GRPO loop adopting CUDA-Agent (arXiv 2602.24286): the model writes a solution, the forger-bench grader runs+verifies+ measures real server metrics, and a discrete milestone reward (-1 incorrect/scaleBug, 1 wasteful, 2 beats-naive, 3 near-optimal) drives RL.
Contamination control
Never trained on a sealed test task; held-out concepts measure optimization skill vs template memorization.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Install Unsloth Studio (macOS, Linux, WSL)
# Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for squaredcuber/forge-optimizer-qwen3.6-35b-a3b to start chatting