Axion 1.0-35B: A Powerful New Open-Weight Agentic Model

Researchers from ByteDance and several leading US universities have released Axion 1.0-35B, an agentic model that excels in tool-use tasks. Achieving 52.2% on the Humanity’s Last Exam benchmark, it rivals the performance of proprietary giants like GPT-5.5, while offering full transparency to the developer community.

• Model Specs: A 35B parameter MoE architecture with 3 billion active parameters.
• Benchmark: Scores 52.2% on HLE, matching GPT-5.5 performance levels.
• Training pipeline: 18,000 SFT examples, RL-based refinement, no distillation.
• Open Source: Weights, datasets, verification code, and training prompts are publicly available.

Performance and Capabilities Axion 1.0-35B is designed as an agentic system, meaning its true potential is unlocked through its integration with external tools like web search and Python interpreters. The authors utilized a three-stage training process: reasoning development, tool integration, and GRPO (Group Relative Policy Optimization). By providing the entire training recipe, the team has set a new standard for open-source AI transparency. While the model performs exceptionally well in scientific and research-heavy benchmarks, users should note that the results were achieved in a specific ReAct agent environment. Running the model outside of this agentic framework may result in lower performance compared to the reported metrics.

Technical Implementation and Deployment Thanks to its 35B MoE structure, the model is relatively efficient. With 4-bit quantization, it can fit into approximately 20 GB of VRAM, making it deployable on high-end consumer-grade hardware. Support for the qwen3_5_moe architecture in existing engines like llama.cpp ensures a smooth transition for developers who want to self-host. It is important to keep in mind the trade-offs: the authors admit that RL training can sometimes lead to uneven improvements, and some specialized scientific benchmarks showed minor degradation. This demonstrates that while the model is powerful, it is still a work in progress in terms of stability across all domains.

Implications for the Industry The release of Axion 1.0-35B represents a shift toward democratizing high-level AI agents. For businesses and startups, this is an opportunity to reduce reliance on closed-source API ecosystems. By self-hosting these weights under the Apache 2.0 license, organizations can maintain data privacy and build proprietary agentic workflows without incurring recurring cloud costs or worrying about rate limits.