Editorial illustration for NVIDIA's Star Elastic bundles 30B, 23B, 12B models; 23B hits 85.63 on AIME-2025
NVIDIA's Star Elastic bundles 30B, 23B, 12B models; 23B...
Most new AI models are just old ones with bigger numbers. NVIDIA's latest is the opposite: three models stuffed into one, and the middle one wins.
Star Elastic bundles a 30-billion parameter model, a 23-billion, and a 12-billion into a single checkpoint. You don't train three separate models. You train one, then slice it.
The 23B version scores 85.63 on the AIME-2025 benchmark. That beats the Qwen3-30B-A3B's 80.00. It does more with less.
The efficiency gains are obscene. Training costs drop by a factor of 360 compared to building each model from nothing. Even against the best previous compression tech, which needed a separate distillation run for every model size, this method is seven times better. The practical upside is speed: the 12B variant runs 2.4 times faster than its 30B parent on an H100 GPU.
The Elastic-23B notably scores 85.63 on AIME-2025 versus Qwen3-30B-A3B’s 80.00, despite having fewer active parameters.
This isn't a marginal improvement. It's a different way of thinking. For years, the rule was simple: bigger model, better performance, higher cost.
Star Elastic breaks that by letting you pick your size from a single trained set of weights. The frontier is no longer just raw scale. It's flexibility.
The next wave of practical AI won't be about who has the most parameters, but who can do the most with them.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications
From Archives Understanding 'Compute': The Core Power Driving Modern AI Models
From Archives Fields Medalist: ChatGPT 5.5 Pro produced PhD-level math proof in under an hour
From Archives NVIDIA and ServiceNow launch autonomous AI agents using Blackwell
Business & Startups
From Archives NVIDIA Magpie TTS models power agentic AI pipeline from cloud to car
Industry Applications
Common Questions Answered
What is unique about NVIDIA's Star Elastic model architecture compared to traditional AI models?
Star Elastic bundles three different model sizes (30B, 23B, and 12B parameters) into a single checkpoint rather than training separate models independently. Instead of training three distinct models, developers train one model and then slice it into different sizes, allowing users to choose their preferred model size from the same trained weights.
How does the 23B model in Star Elastic perform on benchmarks like AIME-2025?
The 23B version of Star Elastic achieves a score of 85.63 on AIME-2025, demonstrating strong performance for its parameter size. This performance is particularly notable because it comes from a model that was derived from a single training process rather than being independently optimized.
Why does Star Elastic represent a shift away from the traditional 'bigger model, better performance' paradigm?
Star Elastic breaks the conventional rule that larger models always perform better at higher costs by enabling users to select their preferred model size from a single trained set of weights. This approach prioritizes flexibility and efficiency over raw scale, allowing developers to balance performance needs with computational constraints without sacrificing quality.
What is the practical advantage of using Star Elastic's single checkpoint approach for deployment?
By training one model and slicing it into multiple sizes, Star Elastic eliminates the need to maintain and deploy three separate trained models, reducing storage and computational overhead. Users can choose the appropriate model size for their specific use case and resource constraints while maintaining consistent quality across all versions from the same checkpoint.
Further Reading
- NVIDIA AI Releases Nemotron-Elastic-12B: A Single AI Model that Gives You 6B, 9B, 12B Variants Without Extra Training Cost — MarkTechPost
- AIME 2025 Benchmark Leaderboard — LLM Stats
- AIME 2025 Benchmark Leaderboard - Artificial Analysis — Artificial Analysis
- Technical Performance | The 2025 AI Index Report — Stanford HAI
