Ox Alpha: A Stealth Model Analysis

The Breakthrough

Ox Alpha is a high-performance, multimodal stealth model currently available for free on OpenRouter that demonstrates architectural and functional parity with GLM 5.3, suggesting it is either a variant of the GLM lineage or a fine-tuned derivative.

What Actually Worked

• Stack Selection: Ox Alpha consistently generated production-ready full-stack applications using Next.js, Drizzle, and Better SQLite, outperforming closed-source models that frequently defaulted to in-memory JavaScript storage.
• Codebase Comparison: Structural analysis revealed that Ox Alpha and GLM 5.3 share rare markers, including identical pass_amount_to_cents functions and specific README formatting, which were absent in other models like DeepSeek V4 Pro or Kimi K3.
• Tokenizer and Encoder Parity: Technical investigation confirmed that Ox Alpha utilizes a tokenizer identical to GLM models and employs a video encoder that uses the same token set as the GLM vision model.
• Performance Benchmarking: While initial marketing claimed superior performance on DeepSWE-bench, independent testing places its capabilities closer to the current generation of open-source models like DeepSeek and Gemini, with a high output capacity averaging 47,000 tokens per task.

Context

Ox Alpha launched as an anonymous model with a 1 million token context window and multimodal capabilities. The author evaluated its performance by prompting it to build a personal finance dashboard and comparing the resulting code structure and UI against several frontier models, including Fable 5, GPT-5.6 Sol, Gemini 3.7 Flash, and GLM 5.3. The analysis suggests that while the model is highly competent and currently free, it is likely a iteration of the GLM architecture rather than an entirely new foundational model.