ByteDance's team has not published just a 35-billion-parameter model. It has made available a complete training recipe: data construction, generation, filtering, verification, prompts, and code. In a field where LLM openness has often stopped at weights, Axion 1.0 tries to make the entire process reproducible.

The path is compact. Three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning. No dedicated mid-training, no direct preference optimization, no on-policy distillation. The first fine-tuning stage builds self-contained reasoning through a model-aware strategy: problems are generated and adversarially strengthened based on solutions the base model already produces. The second stage extends the model to tool-dependent reasoning, with a scalable pipeline of verifiable tasks grounded in real-world data and deterministic computation. Reinforcement learning then strengthens long-horizon reasoning and interaction with external tools and environments.

For AI-RADAR, the point is not the benchmark. An open process changes the equation for local deployment. When a team can reconstruct the training pipeline, audit the steps, and adapt the model to proprietary data, the model stops being an API black box. For organizations with data residency or compliance requirements, this reduces dependence on cloud services and makes the self-hosted option more concrete. It is not an academic detail: the distinction between open weights and an open process changes TCO and risk calculations in regulated contexts.

The documentation does not provide VRAM or throughput specifications, so any hardware evaluation must be done in the field. Still, the 35B scale signals a potentially more accessible local inference profile than models with hundreds of billions of parameters, at equivalent quantization.

There are winners and losers. Labs selling API access to proprietary agentic models face increased competitive pressure: if an open recipe achieves competitive performance on several benchmarks, switching costs for enterprises fall. Teams with existing on-premise infrastructure and MLOps staff can internalize training and fine-tuning without negotiating every update with a vendor. The losers are approaches that bet everything on parameter scale and process secrecy: reproducibility becomes a differentiator.

Structurally, Axion 1.0 signals that openness is shifting from product to method. Publishing weights was the first step; publishing the recipe moves competition to the marginal cost of adaptation. For on-premise deployment, the value is not the single model but the ability to repeat the path on internal data, with deterministic verification and audit. The real test will be whether other groups follow: making transparent not only what was trained, but how.