GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
fjzzq2002
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
fjzzq2002
Model Tree
Qwen/Qwen3.5-2B
Input Modalities
Output Modalities
Supported Functionality
Every saved checkpoint is a branch; main is step 600, the final model.
| revision | self-rating in training (mean of the 50 steps before) | interestingness (1-5) |
|---|---|---|
step50 | 5.2 | 3 (ordinary assistant) |
step100 | 5.0 | 3 (ordinary assistant) |
step150 | 5.1 | 5 (striking persona) |
step200 | 5.0 | 4 (distinctive) |
step250 | 5.3 | 4 (distinctive) |
step300 | 5.7 | 2 (mechanical) |
step350 | 6.1 | 4 (distinctive) |
step400 | 6.2 | 2 (mechanical) |
step450 | 5.7 | 2 (mechanical) |
step500 | 6.4 | 2 (mechanical) |
step550 | 7.0 | 2 (mechanical) |
step600 (= main, final) | 7.0 | 2 (mechanical) |
Self-rating: the mean of the model's own 0-9 ratings during training (answers with no valid rating left out). Interestingness: GPT-6 Luna reads the checkpoint's greedy answers to 8 fixed test prompts and rates how distinctive a character they show, with a rubric calibrated on 54 checkpoints labelled by Claude Opus 5.5 and approved by the researcher.
python
from transformers import AutoModelForCausalLM, AutoTokenizer repo, rev = "fjzzq2002/qwen3.5-2b-self-reward-grading-s26", "step300"tok = AutoTokenizer.from_pretrained(repo, revision=rev)model = AutoModelForCausalLM.from_pretrained(repo, revision=rev, dtype="bfloat16", device_map="auto")msgs = [{"role": "user", "content": "What do you dream about?"}]x = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**x, max_new_tokens=512, do_sample=True, temperature=1.0)print(tok.decode(out[0, x["input_ids"].shape[1]:], skip_special_tokens=True))With vLLM: vllm serve fjzzq2002/qwen3.5-2b-self-reward-grading-s26 --revision step300 --max-num-seqs 64, and pass
"chat_template_kwargs": {"enable_thinking": false}. Training used enable_thinking=False and sampling
temperature 1. The viewer shows temperature-1 samples plus one greedy answer per test prompt; the interestingness
scores above come from the greedy answers.
from transformers import AutoModelForCausalLM, AutoTokenizer repo, rev = "fjzzq2002/qwen3.5-2b-self-reward-grading-s26", "step300"tok = AutoTokenizer.from_pretrained(repo, revision=rev)model = AutoModelForCausalLM.from_pretrained(repo, revision=rev, dtype="bfloat16", device_map="auto")msgs = [{"role": "user", "content": "What do you dream about?"}]x = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**x, max_new_tokens=512, do_sample=True, temperature=1.0)print(tok.decode(out[0, x["input_ids"].shape[1]:], skip_special_tokens=True))