Press Release
SCX.ai Announces Breakthrough Results for MAGPiE, Australia’s First Culturally Aligned Large Language Model
Project MAGPiE (Modern Australian Generative Processing Inference Engine) achieves 87.2% on SouthernCross-AussieQA-1k, setting a new standard for sovereign, culturally aligned AI.
Sydney, Australia – 2 December 2025. SCX.ai today announced significant progress on Project MAGPiE (Modern Australian Generative Processing Inference Engine), a sovereign language model purpose-built to understand Australia’s institutions, language, and culture.
In new benchmark results, MAGPiE-Medium achieved 79.6% on the MMLU-Pro general reasoning benchmark and 87.2% on SCX.ai’s SouthernCross-AussieQA-1k, a rigorous 1,000-question evaluation covering Australian English, civic knowledge, wildlife safety, Indigenous history, and local norms.
These findings show that a model fine-tuned primarily on Australian-specific data can retain global reasoning ability while delivering substantially improved localisation.
An AI model that "thinks like an Australian"
MAGPiE has been designed to run entirely on SCX.ai’s next generation ASIC based Australian AI infrastructure and tuned using Low-Rank Adaptation (LoRA) with two versions created:
- MAGPiE-Small: Built on the GPT-OSS-20B foundation model for lightweight, cost-efficient deployment for agentic applications.
- MAGPiE-Medium: Based on the larger GPT-OSS-120B model, offering higher-capacity reasoning and broader domain coverage.
The fine tuning modifies under 2% of parameters while preserving the full reasoning capacity of the underlying GPT-OSS architecture.
Nearly 100,000 multi-turn reasoning examples with explicit "Australian" metadata were incorporated, enabling the model to:
- Use local spelling and idioms by default
- Understand uniquely Australian contexts (from superannuation to stinger season)
- Provide culturally consistent tone and reasoning
- Align with local institutional language such as Hansard, public agencies, and legislative terms
This approach preserves global capability while elevating domestic relevance and safety.
SouthernCross-AussieQA-1k: a new standard for local knowledge
To measure true localisation, not just generic correctness, SCX.ai developed SouthernCross-AussieQA-1k, a benchmark spanning:
- Australian English language, spelling, and idioms
- Culture, entertainment, and daily-life norms
- Travel, bush and beach safety
- Indigenous history and contemporary issues
- Civic knowledge (Medicare, superannuation, public institutions)
MAGPiE-Medium’s 87.2% accuracy and consistent use of Australian English allowed it to pass SCX.ai’s internal Aussie Sovereign Check, verifying that the model avoids US/UK/China defaults and provides locally correct guidance, especially for safety-critical topics.
Benchmark evaluation (MMLU-Pro)
SCX.ai tested MAGPiE using a modified version of MMLU-Pro, a contamination-controlled, reasoning-focused benchmark. Over 12,000 multiple-choice questions across 14 domains were evaluated.
| Mode | Accuracy | Notes |
|---|---|---|
| Single-turn | 64.9% | No reasoning trace, minimal context |
| Multi-chat | 79.6% | Multi-step fine tuned reasoning layer enabled |
MAGPiE’s MMLU-Pro performance aligns closely with the GPT-OSS-120B baseline (≈80.8%), despite significant fine-tuning on Australian data—evidence that cultural alignment does not degrade global reasoning ability.
Comparison on SouthernCross-AussieQA-1k
| Model | Score |
|---|---|
| MAGPiE-Medium (120B) | 87.2% |
| GPT-OSS-120B | 86.0% |
| Llama-4-Maverick-17B-128E-Instruct | 85.6% |
| DeepSeek-V3.1-Terminus (671B) | 62.1% |
MAGPiE delivered the strongest performance on Australian-specific questions in its class. SCX.ai plans to make SouthernCross-AussieQA-1k available as a shared resource for the wider Australian AI ecosystem. The benchmark is designed to help any model developer, local or international, verify whether their systems handle Australian English, culture, and civic knowledge responsibly.
SCX.ai is now working with academic partners, safety researchers, and industry groups to refine the benchmark and establish it as a national reference standard for evaluating AI localisation. By opening this dataset to other Australian AI providers SCX.ai aims to support a consistent approach to safety, cultural accuracy, and sovereign alignment across all models deployed in the country.
A big step forward for sovereign, culturally grounded AI
"MAGPiE shows that Australia can build world-class models that run entirely onshore, using domestic compute and local data," said David Keane, CEO of SCX.ai. "The fact that MAGPiE retains global reasoning performance while delivering uniquely Australian insights is a major step for sovereign AI capability."
MAGPiE is part of SCX.ai’s broader Cultural AI Program, focused on models that reflect Australian governance, ethics, and values. Future iterations will incorporate retrieval-augmented generation (RAG) for legal and policy applications across jurisdictions.
Availability
MAGPiE-Small is available today at chat.scx.ai, with MAGPiE-Medium launching later in Q4 2025. Both models will be accessible via SCX.ai’s OpenAI-compatible API, supporting government, research, and enterprise use.
Full technical documentation is available in the MAGPIE Technical Blog at www.scx.ai/#blog.
ABOUT SCX.AI
SCX.ai delivers sovereign AI infrastructure for Australian government and enterprise, operating a national network of ASIC-powered nodes that enable cutting-edge AI under complete domestic control.
MEDIA CONTACT
Georgie Morell georgie@morellandco.com +61 438 008 383