GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
MergenAI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
MergenAI
Model Tree
Input Modalities
Output Modalities
Supported Functionality
Wahoo 1.0 Raw was trained using Supervised Fine-Tuning (SFT) on a curated dataset of 1.2 million high-quality examples. The full training run was completed in 2 days on 2×NVIDIA H200 GPUs, pairing efficient large-scale compute with a carefully filtered, Azerbaijani-first instruction corpus.
Training was accelerated end-to-end with Flash Attention 3, enabling significantly higher throughput and memory efficiency across the H200 cluster and allowing longer context windows to be processed without compromise.
To maximize learning quality, the run employed a two-phase Curriculum-Weighted SFT strategy:
This combination produced a cleaner, more robust base checkpoint than uniform single-phase fine-tuning, while keeping the total training budget to just 2 days.
| Property | Value |
|---|---|
| Training method | Supervised Fine-Tuning (SFT) |
| Dataset size | 1,200,000 examples |
| Hardware | 2 × NVIDIA H200 |
| Training duration | 2 days |
| Acceleration | Flash Attention 3 |
| Optimization | Curriculum-Weighted SFT + Selective Loss Masking (2nd half) |
| Primary language | Azerbaijani |
| Ecosystem | HAL-X |
Wahoo 1.0 Raw is designed as a foundational base for Azerbaijani-language assistants, chat applications, content generation, and instruction-following tasks. As a "raw" checkpoint, it is also well suited for further alignment, domain adaptation, and downstream fine-tuning within the HAL-X Ecosystem.
If you use Wahoo 1.0 Raw, please reference the HAL-X Ecosystem.
Built with ❤️ by the HAL-X team.