Beta
Explore
Marketplace
Neural Labs
Chat
Wallet
Docs
Qwen3-4B-Instruct-2507-ja-4bit – AI Model by taniguchi-kyoichi | AlphaNeural AI
You can deploy this model and start earning money today!
taniguchi-kyoichi
/
Qwen3-4B-Instruct-2507-ja-4bit
like
0
mlx
safetensors
qwen3
4bit
lora
japanese
on-device
ja
en
Qwen/Qwen3-4B-Instruct-2507
adapter
apache-2.0
4-bit
us
Views
No views yet
Model card
Files and Versions
Community
API
Deploy
Qwen3-4B-Instruct-2507-ja-4bit
Japanese fine-tuned version of Qwen3-4B-Instruct-2507, quantized to 4-bit for on-device inference.
Model Details
Base Model
:
Qwen3-4B-Instruct-2507
Fine-tuning
: QLoRA (rank 8, 16 layers, 1000 iterations)
Training Data
:
kunishou/databricks-dolly-15k-ja
Quantization
: 4-bit (MLX format)
Model Size
: ~2.1GB
Peak Memory
: ~2.3GB
Intended Use
On-device Japanese language model for iPhone (3GB+ available memory) and Apple Silicon Macs. Optimized for Japanese instruction-following tasks.
Usage with MLX
Training Details
Framework
: MLX-LM 0.30.7
Method
: QLoRA on 4-bit quantized base
LoRA Config
: rank=8, alpha=16, 16 layers
Learning Rate
: 1e-5
Batch Size
: 1
Iterations
: 1000
Hardware
: Apple M3 16GB
Final Val Loss
: 1.743
Best Val Loss
: 1.537 (iter 300)
