Ox Alpha's video encoder spends token-for-token identical budgets to GLM-5V-Turbo on four controlled test videos — matching frame-sampling rate, duration scaling, and resolution scaling. Combined with a shared tokenizer artifact, matching house style, Zhipu's prior use of the stealth channel, and a conspicuous product gap (the community's top ask, a unified vision flagship, missing from the text-only GLM-5.3 launch six days earlier), the most likely identity is a unified multimodal GLM — a GLM-5.3V or GLM-5.5.
The model
stealth/ox-alpha appeared on OpenRouter August 20, 2026 — anonymous, free for one week, and the fifth stealth release in a series whose previous four were all claimed by Chinese labs (Zhipu, Xiaomi, Ant Group, Meituan). The operator claims 100T tokens/day serving capacity.
Evidence
1. Encoder fingerprint (strongest)
Identical prompts and media were sent to Ox Alpha and candidate models; usage.prompt_tokens minus a text-only baseline isolates the vision encoder's spend per asset. Video results:
| 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 |
The match covers three independent encoder design choices: fps-invariance (5fps and 30fps tokenize identically, so both sample at a fixed internal rate), duration scaling (~147 tokens/second for both), and per-frame resolution scaling (360p→720p multiplies cost identically). Every other model has a distinct signature. Image-path counts sit within the GLM family's range but are not exact matches — consistent with a newer unified model with a tweaked image patch configuration.
2. Tokenizer artifact
Independent testers report Ox Alpha trips the same "dirty token" that affects MiMo, Qwen, and GLM models — a shared tokenizer-lineage tell. DeepSeek, which uses its own tokenizer, was not flagged.
3. Output style
Measured across agent traces on this machine: Ox Alpha writes emoji-decorated syntheses (section emojis, red/yellow/green triage dots, checkmark lists) at ~1.3 emoji per 1,000 characters, where Claude Fable 5, GPT-5.6 (Codex, 6M chars sampled), and Grok 4.6 all measure ~0. That triage-dot style is characteristic of GLM and Qwen agentic output.
4. Release context
GLM-5.3 shipped text-only on August 14; vision was the community's most-requested feature and Zhipu's flagship line has never unified text and vision. Ox Alpha appeared six days later with exactly that combination. Zhipu has used this stealth channel before (Pony Alpha → GLM-5).
5. Agentic competence
In a real 6-session orchestration run (5 subagents exploring repos), Ox Alpha made 69 tool calls with a single error — a concurrency-cap rejection it recovered from in one step — no retry loops, no duplicated work, and terse reasoning (0.3–3KB of thinking per session). Consistent with the operator's "sustained agentic work" positioning, and now quantified by the DeepSWE run below.
Case by lab
| Lab | For | Against | Read |
|---|---|---|---|
| Zhipu (GLM) | Exact video-encoder token match with GLM-5V-Turbo; tokenizer family; emoji/triage style; stealth-channel precedent; clear product motive after text-only 5.3; DeepSWE step profile closest to GLM-5.3 | 1M context far exceeds GLM-5V-Turbo's 203K; image-path counts not an exact match | Most likely |
| Xiaomi (MiMo) | Tokenizer family; stealth precedent (Hunter/Healer Alpha); MiMo v2.5 lists a 1.05M context adjacent to Ox's | Video encoder signature clearly different (910-token floor per clip, shallow duration scaling); faster decode (~63 tps) | Possible |
| Qwen (Alibaba) | Tokenizer family; mature video multimodal line; 1M contexts standard | Different encoder signature (408/1,288 pattern); just shipped Qwen 3.8 openly; has never used the stealth channel | Unlikely |
| DeepSeek | V4 Pro 0813 shares the exact 220 context; serving-efficiency architecture fits the 100T tokens/day claim | No video capability ever shipped; no vision model on OpenRouter at all; different tokenizer; releases open weights rather than stealth previews; own GA release was 7 days prior | Effectively ruled out |
| Google (Gemini) | 220 context is Google's signature figure; video input; early style impression | Tokenizer artifact places it in the Chinese-lab lineage; all four prior stealth models in this series were Chinese labs | Ruled out |
Size estimate
Total parameter count is not recoverable from the outside, but active parameters can be bounded from sustained decode throughput against known-size models — ideally on the same lab's serving stack, which the attribution makes likely here.
| Model | Throughput | Known size |
|---|---|---|
| Ox Alpha | ~49 tps | unknown |
| GLM-5V-Turbo | ~46 tps | 744B total / 40B active |
| Qwen 3.8 Max | ~43 tps | undisclosed |
| MiMo v2.5 | ~63 tps | undisclosed |
Estimate: roughly 40B active parameters, several hundred billion to ~1T total. Ox Alpha decodes within a few percent of GLM-5V-Turbo (40B active), suggesting a similar active-parameter budget if served on comparable infrastructure. Given the flagship positioning and 5× larger context than GLM-5V-Turbo, total parameters likely sit at or above GLM-5V-Turbo's 744B — in the same class as contemporary flagship MoEs (GLM-5V-Turbo 744B/40B; DeepSeek V4 Pro 1.57T/48B).
Throughput conflates active parameters with hardware, batch load, quantization, and speculative decoding. The claimed 100T tokens/day capacity also implies serving optimizations that could mask a larger model. Treat the active-parameter figure as an order-of-magnitude bound, not a measurement.
DeepSWE benchmark run
We ran Datacurve's DeepSWE (original long-horizon engineering tasks, graded in sandboxed containers) on Ox Alpha using the official harness — Pier + mini-swe-agent, the same agent used for all leaderboard runs — on the deterministic 10-task subset (--n-tasks 10 --sample-seed 0).
Both failures were near-misses, not breakdowns: scc-bounded-memory-spilling passed 28 of 31 new tests with zero regressions; vulture-persistent-analysis-cache passed 23 of 24 new tests but regressed 4 existing ones. The eight passes include tasks graded against large regression suites (51,469 existing tests kept green on the Meriyah parser task; 3,134 on FastAPI).
Context from the public leaderboard (full 113 tasks, so not directly comparable — a 10-task subset carries roughly ±13 points of sampling noise): claude-opus-5 74%, gpt-5.6-sol 73%, claude-fable-5 70%, glm-5.3 69%, gemini-3.7-flash 65%, deepseek-v4-pro 63%, qwen3.8-max 57%. Ox Alpha's 80% on the subset is directionally top-tier.
One more attribution datapoint fell out: Ox Alpha's step profile (117 mean steps) is closest to GLM-5.3's leaderboard profile (124 steps) — a many-small-steps working style — versus GPT-5.6-sol (61), Claude Fable 5 (88), or Claude Opus 5 (99). And it does not show the step-inflation-without-payoff pattern of gemini-3.7-flash (125 steps at 65%) or deepseek-v4-pro (155 steps at 63%): the steps convert to passes.
Method notes
- Test media are deterministic ffmpeg renders (gradient PNGs at four resolutions; testsrc MP4s varying duration, fps, and resolution independently), so per-model differences isolate the encoder.
- Media token cost =
prompt_tokensfor prompt+media minusprompt_tokensfor the identical text-only prompt, per model. - Style measurements come from local agent-session transcripts (pi, Claude Code, Codex CLI, Grok CLI) on the same machine, August 2026.
- DeepSWE run: Pier 0.3.1, mini-swe-agent, local Docker (OrbStack), 4 concurrent trials; per-trial trajectories and
reward.jsongrades inresults/deepswe-jobs/. - Probe code and raw per-model JSON:
~/Developer/sandbox/ox-alpha-probe(probe.mjs,results/). - Public-comms sources: OpenRouter listing and announcement, tester tokenizer analyses, and third-party coverage of the stealth-model series, GLM-5V-Turbo, GLM-5.3, and DeepSeek V4 Pro 0813.