Skip to content
ChristopherOx Alpha
Index
  1. Briefing
  2. You are in

Open file · unclassified

Who Ox Alpha really is

A stealth model dropped on 20 August 2026. Free, a 1-million-token context, it sees text, image and video. Nobody put a logo on it. This atlas gathers the evidence and explains it from zero.

  1. You will understand
  2. Separate what OpenRouter published from what the community infers.
  3. See the model as five boxes, not a rumor.
  4. Know where to start if you do not speak the jargon.

ID

stealth/ox-alpha

Context

1,048,576

Exactly 2²⁰ tokens

Status

Unclaimed

STRONG HYPOTHESIS · UNCONFIRMED

Leading thesis · not a passport

Unpublished multimodal variant of Z.ai / Zhipu AI’s GLM-5.x family (likely GLM-5.3 Flash/Air or GLM-5.3V).

Someone taped over a car logo. The engine, the toll booth, the radio and the dashboard error are Z.ai’s. The public GLM-5.3 car has no rear camera; this one does. So it is not the brochure car — it is the camera prototype they have not unveiled.

Nobody has claimed the model. OpenRouter only routes it. Manifold is a market that moves. A 21 Aug 2026 snapshot was ~85% Z.ai; later quotes have been lower. Neither 85 nor 63 is a passport.

How a model “thinks”, in 5 boxes

Fingerprints hide in each box. Ox Alpha leaves GLM marks in the three middle ones.

Before you touch

If someone quotes Manifold at 63%, which layer is that?

Pick one. Then move the control and compare.

Why

Layer 3. The market is neither the card nor the knife. Turn the mask until the thin wire.
How a model “thinks”, in 5 boxes. Three layers: official card (solid), GLM hypothesis (copper), market rumor (thin). Not a passport.
  1. 01

    Text in

    “Hello world”

  2. 02

    Tokenizer

    cuts into IDs

  3. 03

    System wrapper

    +75 hidden tokens

  4. 04

    Model

    MoE experts

  5. 05

    Reply

    output tokens

Confidence mixer

A rubric, not Bayes. Drag how much you trust each test. The ceiling is locked: no live run and no lab statement means no passport. Manifold is the market, not this bar.

Before you touch

If you raise only the stealth playbook, who should rise more?

Pick one. Then move the control and compare.

Why

Xiaomi rises more on this rubric. The playbook pushes; the tokenizer does not. Lower the playbook and GLM stays near the ceiling.
Confidence mixer. Four rubric bars. The 90% ceiling is locked: no live run and no lab statement.

Z.ai / GLM-5.x

82%

Your rubric · 90% ceiling

Xiaomi MiMo-V3

24%

Playbook helps; the knife does not

Fine-tune / Composer

22%

Tokenizer yes, z.ai kitchen less

Inkling · Mira

6%

Audio ≠ video. Ruled out.

Decision tree (yes / no)

Each question cuts candidates. What remains is an unpublished multimodal GLM.

Decision tree (yes / no). Each yes/no cut removes candidates. What remains is unpublished multimodal GLM — a hypothesis, not a verdict.

1. Tokenizer = GLM?

Yes, Δ=+75 → continue

No → another family

Start here