LLM EXPLORER 66,420 MODELS INDEXED

MiRAI 4B by Keanual

By Keanual · 510 downloads
HF Repo ↗ Copy URL

MiRAI 4B is an open-source language model by Keanual. Features: 4b LLM, VRAM: 8.1GB, Context: 40K, License: apache-2.0.

Base model:finetune:qwen/qwen3...   Base model:qwen/qwen3-4b   Conversational Dataset:freedomintelligence/al... Dataset:freedomintelligence/al...   Dataset:yahma/alpaca-cleaned   En   Endpoints compatible   Id   Indonesian   Ja   Japanese   K-ai   Mirai   Multilingual   Qlora   Qwen3   Region:us   Safetensors   Sharded   Tensorflow   Unsloth
Model Card on HF 🤗: https://huggingface.co/K-Ai-Official/MiRAI-4B 

MiRAI 4B Parameters and Internals

LLM NameMiRAI 4B
Repository 🤗https://huggingface.co/K-Ai-Official/MiRAI-4B 
Base Model(s)  Qwen/Qwen3-4B   Qwen/Qwen3-4B
Model Size4b
Required VRAM8.1 GB
Updated2026-11-02
MaintainerKeanual
Model Typeqwen3
Model Files  5.0 GB: 1-of-2   3.1 GB: 2-of-2
Supported Languagesen id ja
Model ArchitectureQwen3ForCausalLM
Licenseapache-2.0
Context Length40960
Model Max Length40960
Transformers Version5.17.0
Tokenizer ClassQwen2Tokenizer
Padding Token<|vision_pad|>
Vocabulary Size151936
Torch Data Typebfloat16
Errorsreplace

Best Alternatives to MiRAI 4B

Best Alternatives
Context / RAM
Downloads
Likes
Qwen3 4B Instruct 2507256K / 8.1 GB3109972911
GRPO 4 70256K / 8.1 GB120
FastContext 1.0 4B SFT256K / 8.1 GB5735357
...2.5 Flash Lite Preview Distill256K / 8.1 GB311
Lightning 4B256K / 8.1 GB136
Qwen3 4B Thinking 2507256K / 8.1 GB291023606
SRA RiskGate 4B256K / 8 GB26621
Dwight 4B256K / 8.1 GB37410
Qwen3 4B Instruct 2507 FP8256K / 5.2 GB110074478
...orm3200 Impl Verl S11 Step 039256K / 8.1 GB5801
Note: green Score (e.g. "73.2") means that the model is better than K-Ai-Official/MiRAI-4B.