Downloads · 30 days
0
Llama-Sahabat-AI-v2-70B-IT-GGUF
blackshell69/Llama-Sahabat-AI-v2-70B-IT-GGUF
Llama-Sahabat-AI-v2-70B-IT-GGUF is a machine learning model from blackshell69. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as llama3.
GGUF quantizations of Sahabat-AI/Llama-Sahabat-AI-v2-70B-IT, quantized locally using llama.cpp.
Access
Public
Updated Jun 17, 2026
Repo size
—
Likes
0
Public
Repo makeup
Click a slice to open those files.
.md2.5 KB · 62%
At a glance
- License
- llama3
- Access
- Public
- Created
- Jun 17, 2026
- Updated
- Jun 17, 2026
- SHA
- 146e0fb5
Base models
Model card
From the Hugging Face model README
Llama-Sahabat-AI-v2-70B-IT — GGUF Quantizations
GGUF quantizations of Sahabat-AI/Llama-Sahabat-AI-v2-70B-IT, quantized locally using llama.cpp.
About the base model
Sahabat-AI adalah model bahasa besar (LLM) yang dikembangkan secara kolaboratif oleh BRIN, GoTo, dan Bukalapak untuk mendukung ekosistem AI berbahasa Indonesia. Model ini dilatih menggunakan data bahasa Indonesia yang kaya dan beragam, sehingga mampu memahami konteks budaya, bahasa, dan kebutuhan spesifik pengguna Indonesia dengan lebih baik.
Sahabat AI is a Large Language Model (LLM) collaboratively developed by BRIN (National Research and Innovation Agency), GoTo, and Bukalapak to support the Indonesian-language AI ecosystem. Trained on rich and diverse Indonesian language data, it better understands the cultural context, language nuances, and specific needs of Indonesian users.
Available quantizations
| Repo | Quantization | Size | BPW | imatrix | Best for |
|---|---|---|---|---|---|
| blackshell69/Llama-Sahabat-AI-v2-70B-IT-Q8_0 | Q8_0 | 70 GB | 8.50 | No | Near-lossless, 80+ GB VRAM |
| blackshell69/Llama-Sahabat-AI-v2-70B-IT-Q3_K_S | Q3_K_S | 29 GB | 3.50 | No | Best quality that fits a 32 GB GPU |
| blackshell69/Llama-Sahabat-AI-v2-70B-IT-Q2_K_S | Q2_K_S | 23 GB | 2.77 | Yes | Smallest footprint, quality trade-off |
Quantization notes
- Q8_0 — converted directly from bf16 safetensors; effectively lossless
- Q3_K_S — good quality-size trade-off; fits comfortably on a single 32 GB V100
- Q2_K_S — uses an importance matrix (imatrix) generated from Q3_K_S + groups_merged.txt calibration data for better weight selection at extreme compression
Usage
Load with any llama.cpp-compatible runner:
llama-cli -m llama-sahabat-70b-Q3_K_S.gguf -p "Halo, apa kabar?" -ngl 40
Hardware requirements
| Quantization | Min VRAM (full offload) | CPU RAM (no GPU) |
|---|---|---|
| Q2_K_S | ~25 GB | ~28 GB |
| Q3_K_S | ~32 GB | ~34 GB |
| Q8_0 | ~75 GB | ~78 GB |
- License
- llama3
- Languages
- id, en
- Created
- Jun 17, 2026
- Updated
- Jun 17, 2026