GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
VinayRevanuru
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
VinayRevanuru
Model Tree
Qwen/Qwen2.5-0.5B
Input Modalities
Output Modalities
Supported Functionality
A Qwen2 architecture reconfigured from scratch — not a fine-tune. It takes the tokenizer (and the idea) of Qwen/Qwen2.5-0.5B, then shrinks the model down to fit a modest GPU and a modest evening:
Qwen2Config built from scratch with the official Qwen2.5-0.5B tokenizer
so the small model ships with the same vocabulary (151,936 tokens).Salesforce/wikitext (wikitext-2-raw-v1), empty lines filtered
out, sequences capped at 512 tokens.Trainer on the GTX 1060 Mobile: batch size 2,
gradient accumulation 32 (effective batch 64), gradient checkpointing on,
8-bit AdamW (adamw_bnb_8bit) just to keep the optimizer from eating
the whole 6 GB, fp16/bf16 disabled — Pascal does its best but half precision
is not its friend.Looking at loss curves and smiling. Testing memory tricks on old hardware. Proof that "AI winter in your bedroom" doesn't have to mean a cluster. This model will say confident nonsense and occasionally something that reads like a sentence, and that's exactly the point.
Production anything. Factual questions. Opinions. Your homework. It has never read a book longer than 512 tokens.
python
from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("VinayRevanuru/qwen2-100m-wikitext")model = AutoModelForCausalLM.from_pretrained("VinayRevanuru/qwen2-100m-wikitext") inputs = tok("Once upon a time", return_tensors="pt")out = model.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.8)print(tok.decode(out[0], skip_special_tokens=True))To the GTX 1060 for not catching fire, and to HuggingFace for making from- scratch training feel like something a person can still do at their kitchen table.