GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Liquor-06
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Liquor-06
Model Tree
Qwen/Qwen3-1.7B-Base
Input Modalities
Output Modalities
Supported Functionality
The model answers in the format it was trained on: reasoning inside <think>...</think>, then a
final line ANSWER: <number>. Ask in the same way:
python
from transformers import AutoModelForCausalLM, AutoTokenizer name = "Liquor-06/Qwen3-1.7B-Base-GSM8K-distillops"tok = AutoTokenizer.from_pretrained(name)model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto") question = "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"suffix = "\n\nSolve step by step. Give the final answer on the last line as: ANSWER: <number>"inputs = tok.apply_chat_template([{"role": "user", "content": question + suffix}], add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))Generation stops at <|im_end|>: it is listed with <|endoftext|> in generation_config.json, which
transformers and vLLM both read (vllm serve Liquor-06/Qwen3-1.7B-Base-GSM8K-distillops).
deepseek-flash, thinking off, temperature 0.7) answered the 7,473 GSM8K
training problems; an answer was kept when its final number matched the gold answer and it stayed
under 250 words, with up to 3 attempts per problem. 7,274 problems (97.3%) were kept:
Liquor-06/gsm8k-deepseek-verified.Commands, configs and the full write-up: github.com/Zihao-Wu06/distillops.
Trained and evaluated only on GSM8K-style word problems in English; it has not been tested on harder math or other tasks and is not a general assistant. Training is not bit-for-bit repeatable on GPUs: an earlier run of the same recipe scored 80.4%.
Apache-2.0, as for the base model. GSM8K is MIT-licensed; the training answers came from the DeepSeek API (check DeepSeek's terms of use for your use case).
from transformers import AutoModelForCausalLM, AutoTokenizer name = "Liquor-06/Qwen3-1.7B-Base-GSM8K-distillops"tok = AutoTokenizer.from_pretrained(name)model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto") question = "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"suffix = "\n\nSolve step by step. Give the final answer on the last line as: ANSWER: <number>"inputs = tok.apply_chat_template([{"role": "user", "content": question + suffix}], add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))