Axion 1.0 Hits Open Source: 3B Active Parameters Matching GPT-5.5

Researchers from ByteDance, Duke University, and the University of Maryland released Axion 1.0, an open MoE agent that scored 52.2% on Humanity’s Last Exam. That ties GPT-5.5 on the exact same benchmark while leaving Gemini 3.1 Pro behind in the dust.

The underlying architecture relies on Qwen with a total footprint of 35 billion parameters, though only three billion activate per token during inference. It runs on a frugal diet of 18,000 supervised fine-tuning examples followed by reinforcement learning, dodging standard distillation shortcuts like a budget-conscious student skipping pricey meal plans.

Unlike typical corporate whitepapers that guard their training recipes like state secrets, the team published the full dataset, system prompts, verification code, and an Apache 2.0 checkpoint on Hugging Face. Supervised tuning of the 35B model took a mere 90 GPU hours on an H100, while the heavy lifting went into a three-stage reinforcement learning loop powered by Python and web-browsing tool calls.

Community hackers already packed the weights into a 4-bit GGUF build that slips into 20 GB of VRAM. The catch lies in the tooling setup, as the benchmark victory heavily relies on a custom agentic loop executing code, leaving pure standalone text generation with noticeably more modest gains.

Big tech spent billions convincing the industry that frontier intelligence requires a private datacenter the size of a municipal power plant. Now a lean model running on consumer silicon and open code just proved that clever reinforcement pipelines make raw scale look like an expensive illusion.

Source: arXiv