GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ubermenchh
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
ubermenchh
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
Held-out secrets (never a secret in SFT or RL), 400 games. A game starts with 0 to 4 prior guesses and has 6 turns in total. Temperature 1.0, 256 tokens at most, thinking off.
| Player | Wins | Mean reward | Games ended on an illegal reply |
|---|---|---|---|
| Stock Qwen3-1.7B (100 games, train split) | 0% | 0.001 | 98% |
| After SFT | 7.5% | 0.107 | 69% |
| After SFT + RL, run 1 (this model) | 13% (95% interval 10% to 17%) | 0.1955 | 38.5% |
| After SFT + RL, run 2 (same setup, other seed, not published) | 9.75% (7% to 13%) | 0.1596 | 42% |
| Random legal player | 0% | 0.100 | 0% |
| Solver (random word that fits all feedback) | 99% | 1.120 | 0% |
This model is run 1, the better of two runs, so 13% is a high pick. A fair summary: 80 steps of RL give 10% to 13% wins and about 40% illegal replies, from 7.5% and 69% after SFT. The gain in wins is small and varies between seeds; the drop in illegal replies repeated in both runs.
An illegal reply (not a word, a repeat, wrong length) ends the game. Reward is not the win rate. The model still makes mistakes that a solver would not make: it can name a word that breaks the constraints it just wrote. The best published result I know is 14.5% on a 7B, with a different measure, so the two numbers are not equal.
Use thinking off. The model reads a system prompt, then the board and the known constraints
(letters spaced, symbols = right place, - wrong place, . absent), and replies with a short
line and <guess>WORD</guess>.
python
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "ubermenchh/qwen3-1.7b-wordle-rl"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto") SYSTEM = """You are playing Wordle. Guess the secret 5-letter word. After each guess you get one symbol per letter: = correct letter, correct position - letter is in the word, wrong position . letter is not in the word Think briefly, then answer with exactly: <guess>WORD</guess>Never repeat a word you have already guessed.""" user = """ S L A T E | . . = . = slot 1: ? slot 2: ? slot 3: A slot 4: ? slot 5: EAbsent: L S TGuessed: SLATE Guesses remaining: 5Make your next guess.""" text = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}], tokenize=False, add_generation_prompt=True, enable_thinking=False,)out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=64, do_sample=True)print(tok.decode(out[0], skip_special_tokens=True).split("assistant")[-1])The environment, reward and training configs are in the project repository.
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "ubermenchh/qwen3-1.7b-wordle-rl"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto") SYSTEM = """You are playing Wordle. Guess the secret 5-letter word. After each guess you get one symbol per letter: = correct letter, correct position - letter is in the word, wrong position . letter is not in the word Think briefly, then answer with exactly: <guess>WORD</guess>Never repeat a word you have already guessed.""" user = """ S L A T E | . . = . = slot 1: ? slot 2: ? slot 3: A slot 4: ? slot 5: EAbsent: L S TGuessed: SLATE Guesses remaining: 5Make your next guess.""" text = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}], tokenize=False, add_generation_prompt=True, enable_thinking=False,)out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=64, do_sample=True)print(tok.decode(out[0], skip_special_tokens=True).split("assistant")[-1])