Qwen3-4B-Instruct-2507 is an updated version of the Qwen3-4B non-thinking mode language model with 4 billion parameters and 262,144 tokens of native context length. The model features significant improvements across general capabilities including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage, along with substantially enhanced long-tail knowledge coverage across multiple languages and better alignment with user preferences for subjective and open-ended tasks. The model operates as a causal language model trained through pretraining and post-training stages, using 36 layers with grouped query attention. It has been optimized for diverse applications from knowledge-based tasks to reasoning, coding, creative writing, and agentic use cases with tool calling capabilities.
Alibaba
available local models on Mirai:
available local models on Mirai:
Name
Quantisation
Size
Qwen2.5-Coder-0.5B-Instruct
No
0.5B
Quant.
No
Size
0.5B
Qwen2.5-Coder-1.5B-Instruct
No
1.5B
Quant.
No
Size
1.5B
Qwen2.5-Coder-14B-Instruct
No
14B
Quant.
No
Size
14B
Qwen2.5-Coder-32B-Instruct
No
32B
Quant.
No
Size
32B
Qwen2.5-Coder-3B-Instruct
No
3B
Quant.
No
Size
3B
Qwen2.5-Coder-7B-Instruct
No
7B
Quant.
No
Size
7B
Qwen3-0.6B
No
0.6B
Quant.
No
Size
0.6B
Qwen3-0.6B-MLX-4bit
No
0.6B
Quant.
No
Size
0.6B
Qwen3-0.6B-MLX-8bit
No
0.6B
Quant.
No
Size
0.6B
Qwen3-1.7B
No
1.7B
Quant.
No
Size
1.7B
Qwen3-1.7B-MLX-4bit
No
1.7B
Quant.
No
Size
1.7B
Qwen3-1.7B-MLX-8bit
No
1.7B
Quant.
No
Size
1.7B
Qwen3-14B
No
14B
Quant.
No
Size
14B
Qwen3-14B-AWQ
No
14B
Quant.
No
Size
14B
Qwen3-14B-MLX-4bit
No
14B
Quant.
No
Size
14B
Qwen3-14B-MLX-8bit
No
14B
Quant.
No
Size
14B
Qwen3-32B
No
32B
Quant.
No
Size
32B
Qwen3-32B-AWQ
No
32B
Quant.
No
Size
32B
Qwen3-32B-MLX-4bit
No
32B
Quant.
No
Size
32B
Qwen3-4B
No
4B
Quant.
No
Size
4B
Qwen3-4B-Instruct-2507 is an updated version of the Qwen3-4B non-thinking mode language model with 4 billion parameters and 262,144 tokens of native context length. The model features significant improvements across general capabilities including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage, along with substantially enhanced long-tail knowledge coverage across multiple languages and better alignment with user preferences for subjective and open-ended tasks. The model operates as a causal language model trained through pretraining and post-training stages, using 36 layers with grouped query attention. It has been optimized for diverse applications from knowledge-based tasks to reasoning, coding, creative writing, and agentic use cases with tool calling capabilities.
Alibaba
available local models on Mirai:
Name
Quantisation
Size
Qwen2.5-Coder-0.5B-Instruct
No
0.5B
Quant.
No
Size
0.5B
Qwen2.5-Coder-1.5B-Instruct
No
1.5B
Quant.
No
Size
1.5B
Qwen2.5-Coder-14B-Instruct
No
14B
Quant.
No
Size
14B
Qwen2.5-Coder-32B-Instruct
No
32B
Quant.
No
Size
32B
Qwen2.5-Coder-3B-Instruct
No
3B
Quant.
No
Size
3B
Qwen2.5-Coder-7B-Instruct
No
7B
Quant.
No
Size
7B
Qwen3-0.6B
No
0.6B
Quant.
No
Size
0.6B
Qwen3-0.6B-MLX-4bit
No
0.6B
Quant.
No
Size
0.6B
Qwen3-0.6B-MLX-8bit
No
0.6B
Quant.
No
Size
0.6B
Qwen3-1.7B
No
1.7B
Quant.
No
Size
1.7B
Qwen3-1.7B-MLX-4bit
No
1.7B
Quant.
No
Size
1.7B
Qwen3-1.7B-MLX-8bit
No
1.7B
Quant.
No
Size
1.7B
Qwen3-14B
No
14B
Quant.
No
Size
14B
Qwen3-14B-AWQ
No
14B
Quant.
No
Size
14B
Qwen3-14B-MLX-4bit
No
14B
Quant.
No
Size
14B
Qwen3-14B-MLX-8bit
No
14B
Quant.
No
Size
14B
Qwen3-32B
No
32B
Quant.
No
Size
32B
Qwen3-32B-AWQ
No
32B
Quant.
No
Size
32B
Qwen3-32B-MLX-4bit
No
32B
Quant.
No
Size
32B
Qwen3-4B
No
4B
Quant.
No
Size
4B