QuanLLM-Coder-V2.3-14B

1. Introduction

QuanLLM-Coder-V2.3-14B is a coding-enhanced language model obtained by supervised fine-tuning (SFT) of Qwen/Qwen3-14B on 21K execution-verified, decontaminated coding samples. The LoRA adapter has been merged into the base weights, so the model ships as a standard bf16 checkpoint and can be served directly by mainstream inference stacks.

The model retains the thinking-mode behavior of its lineage (long chain-of-thought before the final answer) and is aimed at competitive-programming-style code generation and general coding assistance.

2. Model Summary

Architecture	Causal LM, 40 layers, GQA (40 Q / 8 KV heads)
Total Parameters	14.8B (13.2B non-embedding)
Context Length	32,768 (evaluated at 20,480)
Training Stage	Post-training (SFT) on the instruct base
Weight Format	bf16, ~28GB, standard safetensors

Training recipe

Item	Value
Method	4-bit QLoRA SFT (r=32, alpha=32, all linear layers), merged post-training
Data	KodCode-V1-SFT-R1 (4,200) + nvidia/OpenCodeReasoning split_0 (16,800)
Data hygiene	9-gram decontamination against LiveCodeBench / BigCodeBench; 8,192-token length filter (drop, no truncation)
Config	seq len 8192, packing, effective batch 64, lr 1e-4 cosine, 1 epoch (170 steps)

3. Evaluation

LiveBench 1.0 — Coding (release 2024-11-25), official pipeline, temp 0.6 / top_p 0.95, objective execution-based scoring:

Task	Score
Coding (overall)	52.9
LCB_generation	78.205
coding_completion	27.632

Position on the official coding leaderboard (same release, scores from the official judgment data):

Model	Coding
qwq-32b	72.2
deepseek-r1	67.4
claude-3.5-sonnet-20241022	67.1
qwen2.5-72b-instruct-turbo	56.2
gemini-2.0-flash-001	53.9
QuanLLM-Coder-V2.3-14B (this model)	52.9
deepseek-r1-distill-llama-70b	51.6
gpt-4o-2024-08-06	51.4
gpt-4o-2024-11-20	46.1

At 14B parameters, the model outperforms both GPT-4o releases and DeepSeek-R1-Distill-Llama-70B, and approaches 72B-class open models on coding.

4. Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "meaworm/QuanLLM-Coder-V2.3-14B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "Write a Python function to check whether a string is a palindrome, ignoring case and non-alphabetic characters."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=8192)
print(tokenizer.decode(output[0][inputs.input_ids[0].shape[0]:], skip_special_tokens=True))

24GB GPU inference (vLLM, FP8 dynamic quantization):

from vllm import LLM
llm = LLM(model="meaworm/QuanLLM-Coder-V2.3-14B", quantization="fp8",
          gpu_memory_utilization=0.9, max_model_len=20480)

5. Usage Recommendations

• Sampling: temperature 0.6, top_p 0.95 (the settings used in our evaluation). Greedy decoding also works and slightly reduces variance on short tasks.
• Thinking budget: the model produces long chain-of-thought. For competitive-programming problems, allow at least 16K output tokens to avoid truncation before the final code block.
• Known weakness: code-completion-style prompts (fill-in-the-middle / strict format continuation) are relatively weaker than full-solution generation; prefer asking for complete solutions.

6. License & Attribution

The model weights are released under Apache 2.0, following the base model's license. Training data includes nvidia/OpenCodeReasoning (CC-BY-4.0) — attribution to NVIDIA and the dataset authors is required when redistributing or reporting.

7. Citation

@misc{quanllm-coder-v2.3-14b,
  title  = {QuanLLM-Coder-V2.3-14B},
  author = {QuanLLM Team},
  year   = {2026},
  url    = {https://huggingface.co/meaworm/QuanLLM-Coder-V2.3-14B}
}