Research note: Model specifications and benchmark results are time-bound. Check the dated primary sources below before using them for a technical or purchasing decision.

In mid-2026, an unannounced model codenamed oX Alpha surfaced quietly on open benchmarking leaderboards and OpenRouter. Exhibiting near-perfect 1M-token needle-in-a-haystack retrieval and competitive HumanEval coding reasoning, reverse-engineering its tokenizer pointed to a distilled open-weights architecture based on GLM-5.3.

1. Tokenizer Fingerprinting & Retrieval Benchmarks

Analysis revealed a 150,000-vocabulary BPE tokenizer matching Tsinghua / Zhipu AI's lineage, with custom rotary position embeddings (RoPE) scaled for 1,048,576 tokens without precision loss.

Sources & Further Reading

  • • OpenRouter Model Registry: openrouter.ai — Open weights and API benchmark tracking.
  • • RoPE Scaling Research: Su, J., et al. (2024). RoFormer: Enhanced Transformer with Rotary Position Embedding. Neurocomputing. arXiv:2104.09864.