The mystery model 'ox-alpha' that topped OpenRouter and OpenCode usage charts last week has been revealed as Z.ai's new GLM-5.3-Flash, now live in Kilo across the IDE extension, CLI, and Cloud Agents. The model has 320B total parameters with only 18B active, uses a hybrid linear/sparse attention architecture with an IndexPool optimization, and is priced at roughly a tenth of GLM-5.2 while outperforming it on most benchmarks. It scores near Claude Opus 4.8 on coding and agentic benchmarks (DeepSWE, AutomationBench, Terminal Bench) at a fraction of the cost, and is the first natively multimodal model in the GLM-5 line, with visual self-verification useful for frontend and UI work. Z.ai also disclosed it served the launch on Chinese AI chips with a custom SGLang-based inference stack, claiming efficiency parity with NVIDIA GPUs. Kilo recommends it as a default model for routine coding, parallel agent workflows, code review, and frontend tasks.
Table of contents
What it actually isThe benchmarksVision that’s actually useful for codingOne more thing worth flaggingWhere it fits in your Kilo workflowTry itQuestions this post answers
What model was hiding behind the ox-alpha name on OpenRouter and OpenCode?
Ox-alpha was Z.ai's GLM-5.3-Flash, revealed after a week of topping usage charts under a stealth codename. It has 320B total parameters with 18B active, is the first natively multimodal model in the GLM-5 line, and is priced at roughly a tenth of GLM-5.2 while beating it on most benchmarks. Track new model reveals like this one on daily.dev before deciding what to route your agents to.
How does GLM-5.3-Flash's architecture cut inference cost compared to GLM-5.3?
It combines linear attention for local dependencies with sparse attention for global context via a lightweight indexer, plus an IndexPool technique that compresses four indexer key vectors into one. Against GLM-5.3, this cuts attention compute by 3x and KV cache size by 4.4x, enabling cheap serving even at 1M-token context. Developers evaluating model architectures for cost-sensitive agentic workloads can follow details like this on daily.dev.
How does GLM-5.3-Flash compare to Claude Opus 4.8 on coding benchmarks?
GLM-5.3-Flash scores close to Opus 4.8 across several benchmarks: 84.3 vs 85.0 on Terminal Bench 2.1, 63.4 vs 58.0 on DeepSWE v1.1, and 78.4 vs 76.2 on Toolathlon Verified. On Z.ai's internal Code Bench run through Claude Code, it scores 29.0 versus Opus 4.8's 29.5, at a fraction of the per-task cost (about $0.045 on the Artificial Analysis Intelligence Index). Compare emerging coding models against incumbents like Opus on daily.dev before picking your default.