DALLAS (July 9, 2025). MythWorx has achieved a breakthrough in AI performance, scoring 71.24% accuracy on the rigorous MMLU-Pro benchmark with zero pretraining, no chain-of-thought prompting and no retries. The test spanned more than 12,000 tasks across 14 subjects, each answered in a single attempt. That stands in stark contrast to the methods typically used by large language models (LLMs) and large reasoning models (LRMs).

MythWorx’s 14B-parameter model outperformed models hundreds of times larger, including DeepSeek-R1 (671B) and Llama 4 Behemoth (2T). It also delivered a best-in-class 87.64% accuracy on the math portion, second only to Gemini 2.5 Pro, at a fraction of the size, energy use and compute load.

An architecture built for the real world

Unlike traditional LLMs that require massive datasets and extensive pretraining, MythWorx’s model is built on a hybrid architecture designed for self-improvement and real-time adaptability. Its zero-shot reasoning cuts compute needs by more than 90% compared with traditional models that rely on repeated fine-tuning and prompting, drastically reducing infrastructure costs.

Smarter, smaller, more human

The model’s design introduces dynamic memory, conversational intent alignment and cross-domain adaptability. This allows for:

• Seamless integration into collaborative workflows

• Flexible deployment across secure or low-resource environments

• A compact, license-ready profile for highly specific applications

This isn’t just a model. It’s a reasoning system built to think, adapt and partner with people at a higher level of trust and transparency.

About MMLU-Pro

MMLU-Pro (Massive Multitask Language Understanding Pro) evaluates reasoning and domain generalization across 14 complex subject areas with more than 12,000 questions. It is one of the most rigorous benchmarks for assessing high-level AI performance. More information is available at huggingface.co/datasets/TIGER-Lab/MMLU-Pro.

About MythWorx

MythWorx is pioneering a new class of brain-inspired AI built on human-like reasoning, real-time learning and ultra-low compute. Its proprietary neuromorphic architecture enables systems that understand, adapt and self-improve without massive data, pretraining or energy-intensive infrastructure. Designed for mission-critical enterprise, government and defense applications, MythWorx delivers high-assurance, edge-ready, auditable intelligence that outperforms traditional AI while dramatically reducing cost, latency, hallucinations and power consumption.

Media contact

AJ Jean, PRforMythWorx@bospar.com