Briefing
Stage Test · 2/3

Test

The math, without fear

Why a constant offset is hard evidence, MoE, 2²⁰ and the video toll.

You will understand
Subtract the letterhead by hand.
See MoE and 2²⁰ as a bill and a desk.
Read the kill matrix as text, not only color.

Why a constant is hard evidence

tok_ox(P) − tok_glm(P) = 75    ∀ P

If the knives differed, Δ would jitter with language, code and emoji.
A wrapper (hidden system prompt) adds a FIXED extra. That is the letterhead.

Analogy: two people cut the same newspaper with the same scissors. One always gets 75 extra clippings because they add a letterhead. The scissors are the same.

Lab: subtract the letterhead

Type the token counts two APIs returned for the SAME text. If Δ = 75 every time, the knife is the same.

Before you touch

If GLM returns 12 and Ox returns 87, Δ is...

Pick one. Then move the control and compare.

7599Cannot tell

Why

75. 87 − 12 = 75. That is the letterhead.

GLM-5.3 prompt_tokensOx Alpha prompt_tokens

Δ = tok_ox − tok_glm
Δ = 87 − 12 = 75

H0  same tokenizer + wrapper   →  Δ = 75  for every P
H1  different tokenizer        →  Var(Δ) ≫ 0

Fits the hypothesis: 75 letterhead tokens. If this repeats in English, Chinese, code and emoji, the dictionary is GLM.

Lab: wake the specialists

The building always has 744 “musicians”. You only pay the ones who play. A dense 2T model bills everyone, all the time.

Before you touch

If you wake fewer experts, the bill...

Pick one. Then move the control and compare.

Falls, because only the awake ones are paidRises, because the building is the sameDoes not change: 744B is 744B

Why

Falls. cost ≈ k · P_active · tokens. The whole building is not billed on every word.

744 musicians in the building. You only pay the ones who play. Real GLM-5.3 uses ~40B of 744B.. 744 musicians in the building. You only pay the ones who play. Real GLM-5.3 uses ~40B of 744B.

Active experts per token: 40B

Building (total)

744B

Awake (the bill)

40B

Cheaper than dense 2T

×50.0

Only 5.4% of the building works on each word. Real GLM-5.3 uses ~40B of 744B.

Lab: 2ⁿ is not marketing

GPUs love powers of two. 1,048,576 is not a round “million”. It is exactly 2²⁰.

Before you touch

If n goes up by 1, the cube...

Pick one. Then move the control and compare.

Grows a little, like n+1Doubles (×2)Becomes 20× larger

Why

It doubles. That is why 2¹⁹ to 2²⁰ looks like a desk, not a marketing round number.

Each step multiplies by two. 2²⁰ = 1,048,576, Ox Alpha’s desk and GLM-5.2/5.3’s.. Each step multiplies by two. 2²⁰ = 1,048,576, Ox Alpha’s desk and GLM-5.2/5.3’s.

n = 20 → 1,048,576

2²⁰ = 1,048,576

≈ 10.5 books of 100k tokens. This is Ox Alpha’s desk, and GLM-5.2/5.3’s.

2^10 · 1,024 · a long SMS
2^14 · 16,384 · an essay
2^17 · 131,072 · Ox max output
2^20 · 1,048,576 · Ox / GLM-5.2 context

Video toll booth

Four published clips (Davis / Railway / BohuTANG). Ox and GLM-5V-Turbo charge the same. MiMo and Qwen do not. 5 fps = 30 fps is the trick: the booth bills time, not frames.

Before you touch

If 5 fps bills the same as 30 fps, the cashier is counting...

Pick one. Then move the control and compare.

Raw framesTime (and a specific encoder)Total pixels

Why

Time. The identical Ox = 5V-Turbo toll is the fingerprint; fps invariance is common.

Published numbers, not our own run. Ox and GLM-5V-Turbo charge the same on four clips.. Published numbers, not our own run. Ox and GLM-5V-Turbo charge the same on four clips.

Ox Alpha

296

GLM-5V-Turbo

296

MiMo v2.5

910

Qwen

408

Stamp: SAME CASHIER. 296 = 296 tokens.

Published numbers, not our own run. GLM-4.6V bills differently: “GLM-ish” is not enough.

Kill matrix

Instruments, not vibes. A “kill” cell is not hate: it is a test that suspect cannot explain.

Test	GLM	MiMo	MiniMax	Hy4	Composer	Inkling	West	Seed
Tokenizer +75 / L1 L1=0 vs GLM-5.3. MiMo 176, MiniMax 172, others 148–302. From-scratch does not inherit the knife.
296 toll Ox = GLM-5V-Turbo token-for-token. MiMo 910, Qwen 408. fps-invariance is common; the budget is not.
Audio Ox rejects. MiMo v2.5 accepts. Inkling is audio-native. Softer than 296 (few raw dumps).
1210 + Java PaaS Bilingual 1210 screenshot on Ox. com.wd.paas is Zhipu. 1214-on-Ox is a claim (Pritish), not a photo.
Stealth playbook 4/4 2026 animals = Chinese labs. Xiaomi already burned Hunter. Cursor from-scratch does not use this channel for a flagship.
100T factory Capacity, not a meter. ByteDance already claims 120T/day — best waiter. Does not sign the chef’s passport.

Pick a cell to read the instrument.