GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jhhj25
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
jhhj25
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Given a task description of the kernel's five tuning parameters (RDMA queue pairs, tokens per chunk, dispatch sub-puts, combine streaming batch, communication SMs), the model proposes a configuration as a JSON object. Every reward during training was a hardware measurement: dispatch+combine bandwidth in GB/s on a 16-GPU H200 system, gated by a 16/16 correctness test. No synthetic rewards.
Alternating coordinate ascent between the model and its harness, with the judge frozen:
Five model phases and three harness phases produced this checkpoint. Intermediate checkpoints of every run are published in the companion repos (reef-nccl-ep-qwen3-8b-lora-coevolve-run2/-run4/-run5, -16k-run2, -discovery-16step).
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B")model = PeftModel.from_pretrained(base, "jhhj25/reef-nccl-ep-qwen3-8b")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
Disable thinking mode when sampling configs
(chat_template_kwargs={"enable_thinking": False}), and prompt with a task description
of the parameter space; the model answers with a single JSON config block.