Mystery 'OxAlpha' Model Emerges with Code Ability Rivaling Top Models
机器之心 · wechat · 2026-08-22
Model Performance
- OxAlpha: An anonymous model on OpenRouter featuring 1M context and multimodal input. It achieved an 80% pass rate on DeepSWE (10 tasks), outperforming Fable5Max (65%), GLM-5.3Max (62%), and GPT-5.6SolMax (52%), indicating strong long-context coding potential. Another test subset showed 63%.
- korrine: Appeared on CodeArena with uncertain identity, speculated to be KimiK3.1, Qwen, or MiMo.
Identity Speculation
- OxAlpha likely Zhipu GLM-5.x:
- Visual token consumption matches GLM-5V-Turbo exactly.
- Text tokenizer tests show a fixed offset compared to GLM-5.3.
- Refusal to process audio and inference interface style align with GLM.
- Zhipu has a history of anonymous testing (PonyAlpha).
- Industry Context: Vendors prefer anonymous testing to eliminate brand bias, gather real feedback, and stress-test infrastructure.
Impact
If OxAlpha is indeed a Flash model with this capability, the potential of its full version is highly anticipated.
More from Companies & People
- Local media growth via Meta ads: UGC-style creatives win, exclusion targeting matters — TJLarkin23 · 2026-10-16
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Princeton's Elad Hazan announces new entity for alignment and safety research — HazanPrinceton · 2026-10-16
- Ollama CEO: 8.9M monthly developers, Fortune 500 quietly switching to open models — ollama · 2026-10-16
- Founder: AI reviews contracts in 20 minutes for $5, while big-company legal takes months — bindureddy · 2026-10-16
- Stanford HAI talk: how organizations should measure the value of AI — StanfordHAI · 2026-10-16