flylcw2-2B
A Compact Chinese Language Model
2 min readupdated
The flylcw2-2B model is a text LLM (large language model) with 1.88 billion parameters built on the Qwen3.5 architecture and tokenizer. Trained via supervised instruction tuning, it continues the Seq-Monkey SFT project. The model focuses on the Chinese language and delivers strong results on internal benchmarks (evaluation test suites), outperforming the base version and several comparable alternatives on select tasks. This release suits developers and researchers looking for a compact solution for local deployment.
Key features
- Built on the Qwen3.5 architecture with weights in BF16 SafeTensors format.
- Tested on an internal Chinese benchmark covering Python and SQLite tasks.
- Supports a default context window of up to 8,192 tokens and up to 16,384 tokens maximum.
- Optimized to run on a single graphics card using about 3.79 GiB of VRAM.
What you can do
- Generate Chinese text and answer questions using the model.
- Integrate it into local pipelines or chatbots for automation tasks.
- Test simple Python code generation and SQLite database queries.
Example
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_dir = "./flylcw2-2B"
tokenizer = AutoTokenizer.from_pretrained(model_dir, local_files_only=True)
model = AutoModelForCausalLM.from_pretrained(
model_dir, dtype=torch.bfloat16, local_files_only=True
).cuda().eval()
messages = [{"role": "user", "content": "Return only JSON containing your favorite city and today's weather."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, add_special_tokens=False, return_tensors="pt").to(model.device)
with torch.inference_mode():
tokens = model.generate(**inputs, do_sample=False, max_new_tokens=128, use_cache=True)
print(tokenizer.decode(tokens[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))Read more
DetailsBenchmarks: 110/210 on general tasks, 20/20 on Python, 10/10 on SQLite.
Hardware: trained on 8 × NVIDIA RTX 6000D, VRAM usage during inference is 3.79 GiB.
Limitations: weak performance in arithmetic and math (11/60), potential phrase repetition.










Comments