Ox Alpha: Performance Analysis and GLM Origin Speculation

Performance Benchmarks

Ox Alpha demonstrates frontier-level performance across multiple evaluation suites. On the author's proprietary 'Kingbench', the model achieved a score of 70 out of 80 (87.5%), placing it second only to GLM 5.3. It performed exceptionally well on complex reasoning and technical tasks, securing perfect scores on 3JS contact lens generation, Panda SVG creation, hard math permutations, and Gemma fine-tuning tasks. In a 10-task subset of the DeepSWA benchmark, Ox Alpha achieved an 80% success rate, notably solving the 'Marriott task' in a single attempt, whereas GLM 5.3, GPT 5.6 Soul, and Grok 4.6 failed to solve it in four attempts each.

Evidence for GLM Origin

Technical analysis suggests a high probability that Ox Alpha is a new model from the GLM lab. The primary evidence includes:

• Video Encoder Fingerprinting: The model's tokenization behavior, frame sampling, and duration scaling (147 tokens per second) match GLM 5V Turbo exactly.
• Tokenizer Parity: The model shares an identical vocabulary with GLM 5.3, confirmed by matching token counts across 25 distinct prompts.
• Behavioral Markers: The model utilizes the same emoji-heavy response style characteristic of GLM and Qwen models and rejects audio input, consistent with GLM 5V.
• Exclusionary Testing: Comparative analysis of tokenization and encoding signatures systematically ruled out DeepSeek, Qwen, Xiaomi, and Western-based frontier models.