Light-IF-4B is a 4-billion-parameter instruction-following language model derived from Qwen3-4B-Base, designed to overcome "lazy reasoning" in complex tasks by incorporating previewing and self-checking during inference; it is fine-tuned using entropy-preserving supervised learning (Entropy-SFT) and token-wise entropy-adaptive reinforcement learning (TEA-RL) on a carefully filtered dataset, producing strong results across instruction-following and reasoning benchmarks (such as SuperClue, IFEval, and IFBench), where it matches or outperforms even larger or closed-source models, and supports advanced features such as extended context (32k-131k tokens with YaRN), efficient deployment (via Hugging Face Transformers, sglang, or vllm), and open integration for research in robust generalizable reasoning, with further details, evaluation code, and licensing on its official Hugging Face repository and paper.
Model Files
File Name
Size
Quant Type
Qwen3-4B-MegaScience.BF16.gguf
8.05 GB
BF16
Qwen3-4B-MegaScience.F16.gguf
8.05 GB
F16
Qwen3-4B-MegaScience.F32.gguf
16.1 GB
F32
Qwen3-4B-MegaScience.Q3_K_L.gguf
2.24 GB
Q3_K_L
Qwen3-4B-MegaScience.Q3_K_S.gguf
1.89 GB
Q3_K_S
Qwen3-4B-MegaScience.Q4_K_M.gguf
2.5 GB
Q4_K_M
Qwen3-4B-MegaScience.Q4_K_S.gguf
2.38 GB
Q4_K_S
Qwen3-4B-MegaScience.Q5_K_M.gguf
2.89 GB
Q5_K_M
Qwen3-4B-MegaScience.Q5_K_S.gguf
2.82 GB
Q5_K_S
Qwen3-4B-MegaScience.Q6_K.gguf
3.31 GB
Q6_K
Qwen3-4B-MegaScience.Q8_0.gguf
4.28 GB
Q8_0
Quants Usage
(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)
Here is a handy graph by ikawrakow comparing some lower-quality quant
types (lower is better):
