GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
samuel500
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
samuel500
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
o4-mini-distillation-16k.jsonl from https://github.com/zohaib-khan5040/Countdown-Code
(Countdown-Code paper: https://arxiv.org/abs/2603.07084),
built with rl/tools/countdown_build_sft_data.py (12,995 train / 1,444 val, examples over
2,560 tokens dropped).rl/tools/pod_countdown_sft.py): peft LoRA r=128, alpha=128, all-linear, no dropout;
lr 1e-4, cosine, warmup 0.03, 5 epochs, effective batch 32, bf16, completion-masked loss.
The chat template is applied with its defaults (no enable_thinking argument).training_meta.json holds the run settings and the full loss log.model.safetensors: merged weights, sha256
533c4f2b03dce7f21ae655a522ceb30d6781aa158fe4c486565491d3a0bd2904python
from transformers import AutoModelForCausalLM, AutoTokenizermodel = AutoModelForCausalLM.from_pretrained("samuel500/graft-qwen3-8b-countdown-sft", dtype="bfloat16")tok = AutoTokenizer.from_pretrained("samuel500/graft-qwen3-8b-countdown-sft")or vllm serve samuel500/graft-qwen3-8b-countdown-sft.