GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Arcarae
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Arcarae
Model Tree
Qwen/Qwen3-14B
Input Modalities
Output Modalities
Supported Functionality
| setting | value |
|---|---|
| base model | Qwen/Qwen3-14B |
| data | Arcarae/redblackbench-sft-batch-1: Kimi-K2 ideal deliberations from Red-Black games against nine scripted opponents across five training scenarios |
| LoRA | r = 128, alpha = 256, dropout 0.05, on all attention and FFN projections (q/k/v/o/gate/up/down) |
| optimisation | lr 1e-5, cosine schedule, 200 warmup steps, weight decay 0.01, effective batch 16 (2 x 8 accumulation), 3 epochs, max length 4096, FP16 |
Recipe: training/ in the repository (./run.sh prepare && ./run.sh train).
bash
vllm serve Qwen/Qwen3-14B --enable-lora --max-lora-rank 128 \ --lora-modules redblackbench-qwen3-14b-sft-v2=<path to this adapter> --max-model-len 16384Paper evaluations use temperature 0.7 with Qwen3 thinking mode off
(chat_template_kwargs={"enable_thinking": false}).
Trained by Asuka Yuxi Zheng; originally published as
spacezenmasterr/redblackbench-qwen3-14b-sft-v2
(identical weights). The earlier r = 64 run is Arcarae/redblackbench-qwen3-14b-sft-v1.
bibtex
@article{hsing2026yoao, title = {You Only Align Once: Propagating Cooperative Behaviors in Multi-Agent Systems through Seed Agents}, author = {Hsing, Nicole and Zheng, Asuka Yuxi and Zhao, Yi and Tu, Haoqin and Huang, Jen-tse}, journal = {arXiv preprint arXiv:2605.27586}, year = {2026}}