What Is Ox Alpha?

Best explanation: Zhipu's next-generation unified multimodal GLM GLM-5.3's tokenizer, the GLM-5V line's video encoder, a stronger base than GLM-5.3 — i.e. an unreleased GLM-5.3V / GLM-5.5, stealth-testing at scale before launch. Confidence ~90%.

The model

stealth/ox-alpha — anonymous, appeared on OpenRouter August 20, 2026, free for one week.

1,048,576-token context, 131,072 max output, text + image + video input, reasoning model.

Fifth stealth release in a series; the previous four were all claimed by Chinese labs (Zhipu, Xiaomi, Ant, Meituan).

Operator claims 100T tokens/day serving capacity; measured decode ~49–78 tokens/sec.

Evidence it's Zhipu

Video encoder match (strongest). On four controlled test videos, ox-alpha spends token-for-token identical budgets to GLM-5V-Turbo, matching three independent encoder design choices: fps-invariant frame sampling, ~147 tokens/sec duration scaling, and per-frame resolution scaling. Every other candidate tested (MiMo v2.5, Qwen 3.8 Max, GLM-4.6V) has a clearly different signature.

Test video	Ox Alpha	GLM-5V-Turbo	GLM-4.6V	Qwen 3.8 Max	MiMo v2.5
2s · 30fps · 360p	296	296	1,832	408	910
2s · 5fps · 360p	296	296	1,832	408	910
6s · 30fps · 360p	884	884	5,492	1,288	1,364
2s · 30fps · 720p	1,064	1,064	7,214	1,156	1,134

Text tokenizer match. Independent fingerprinting found ox-alpha's token counts match GLM-5.3 exactly across 25 diverse prompts, modulo a constant +75-token hidden wrapper — requiring an identical vocabulary, not just shared lineage.

Audio discriminator. Ox-alpha rejects audio input with the same routing behavior as GLM-5V (no audio endpoints); MiMo v2.5 — the leading rival theory — accepts and tokenizes audio, a headline V2.5 feature it would have had to drop.

House style. Emoji-decorated syntheses (section emojis, red/yellow/green triage dots, checkmark lists) at ~1.3 emoji per 1,000 chars — the GLM/Qwen signature; Claude, GPT-5.6, and Grok traces measured on the same machine sit at ~0.

Working style. On DeepSWE, ox-alpha averages 117 agent steps per task — nearly identical to GLM-5.3's leaderboard profile (124), unlike GPT-5.6-sol (61) or Claude (88–99).

Motive and precedent. Zhipu shipped GLM-5.3 text-only on August 14; a unified vision flagship was the community's top unmet ask. Zhipu has used this stealth channel before (Pony Alpha → GLM-5).

Size economics. Decode speed within ~6% of GLM-5V-Turbo (744B total / 40B active) implies a ~40B-active MoE — the only size class where a free 100T-tokens/day week is a plausible fleet burn-in rather than fantasy.

Capability check

DeepSWE (official harness, deterministic 10-task subset): 8/10 Pass@1, both misses near-misses.

On those exact tasks, the leaderboard models average: Claude Fable 5 65%, GLM-5.3 62%, Grok 4.6 62%, GPT-5.6-sol 52% (4 attempts each).

It solved meriyah-explicit-resource-declarations single-shot — GLM-5.3, GPT-5.6-sol, and Grok 4.6 went 0/4 on it — while keeping 51,469 regression tests green.

Efficient agentic behavior: in a 69-tool-call orchestration session, one error, zero retry loops, minimal reasoning overhead.

Implication: stronger than GLM-5.3 on identical tasks — consistent with a next-generation checkpoint, not a variant.

Ruled out

Xiaomi MiMo — right size class (V2-Pro: 1T/42B, 1M context) and the same stealth playbook, but the video encoder and audio surface both contradict it.

DeepSeek — no video capability ever shipped, different tokenizer, releases open weights instead of stealth previews.

Google — tokenizer places ox-alpha in the Chinese-lab lineage despite the Gemini-like 2²⁰ context.

Qwen — different encoder signature; just shipped 3.8 openly.

xAI / OpenAI / Anthropic — tokenizer, style, and encoder all mismatch.

Method, briefly

Encoder fingerprint: identical prompt + deterministic ffmpeg test media sent to each model; usage.prompt_tokens minus a text-only baseline isolates the vision encoder's per-asset spend.

Benchmark: Datacurve's Pier + mini-swe-agent (the official leaderboard harness), sandboxed local Docker, --n-tasks 10 --sample-seed 0.