Briefing
Stage Learn · 5/5

Learn

12-question exam

Active recall. Each miss teaches the trap. The sentence you must be able to say out loud.

You will understand
Recall, do not reread.
Each miss teaches the trap.
Leave with a sentence you can say out loud.

1. A token is…

A letter.
Always a whole word.
A scrap of text (subword, sign, chunk of code) the model cuts into.
A model parameter.

Score. A scrap of text (subword, sign, chunk of code) the model cuts into.

“Hi” can be 1. “extraordinarily” can be 3. A line of code, 20. Old SMS counted letters; models bill tokens.

Trap. “A million words” is lazy marketing. 1,048,576 is 2²⁰ tokens, not a round million.

Open

2. OpenRouter is…

The lab that trained Ox Alpha.
The mall / router. It routes. It is not the owner.
Z.ai under another name.
A prediction market.

Score. The mall / router. It routes. It is not the owner.

They write it: “we are not the developer, owner or provider”. The aisle is stealth/ox-alpha. The chef hides.

Trap. “OpenRouter launched Ox” is a lazy headline. Blaming the mall for the passport is error #1.

Open

3. A constant Δ = +75 vs GLM-5.3 means…

Ox is smarter: it uses 75 extra thinking tokens.
Same knife (vocabulary) plus a hidden ~75-token letterhead.
Ox is public GLM-5.3, period, passport.
Ox’s tokenizer is 75× larger.

Score. Same knife (vocabulary) plus a hidden ~75-token letterhead.

If the extra were thinking, Δ would vary. If the knife were other, Δ would not be 75 on EN, DE, CN, code and emoji. Constant = letterhead + same dictionary.

Trap. Same tokenizer ≠ same weights. Zhou: ~95% family, ~70% SKU. Public 5.3 is text-only.

Open

4. The 296/296/884/1064 video toll mostly kills…

Nobody: 5 fps = 30 fps is unique to GLM.
Xiaomi MiMo (910) and Qwen (408). The exact budget is GLM-5V-Turbo.
OpenAI, because they do not accept video.
Public GLM-5.3 itself, which already sees video.

Score. Xiaomi MiMo (910) and Qwen (408). The exact budget is GLM-5V-Turbo.

Qwen and MiMo are fps-invariant too (408=408, 910=910). What kills is the exact number: only 5V-Turbo bills 296. Public GLM-5.3 is text-only.

Trap. Do not inflate: fps-invariance is common. Token-for-token match does not prove Ox IS the public 5V-Turbo endpoint (200K vs 1M).

Open

5. “Ox scores 80% on DeepSWE and beats Fable” is…

The official 113-task score.
A 10-task pop quiz (Davis). The bench author measured ~63%.
Proof it is GPT-5.
The same as GLM-5.3’s official 66.9.

Score. A 10-task pop quiz (Davis). The bench author measured ~63%.

DeepSWE has 113 tasks. 8/10 has CI ~49–94%. winkey_h (author): ~63% subset. Official GLM-5.3: 66.9. Viral 80% flattened the n=10.

Trap. Do not crown a stealth over Fable with ten free throws in the park.

Open

6. Manifold ~80–85% Z.ai is…

A peer-reviewed paper.
This dossier’s rubric.
A prediction market: opinion with skin in the game, not a passport.
A Z.ai statement.

Score. A prediction market: opinion with skin in the game, not a passport.

Who ~80% Z.ai; name GLM-5.3V ~46%. This dossier’s mixer is another bar (90% ceiling). Z.ai has not spoken.

Trap. Memorizing 85% without knowing what Manifold is is the same sin as inflating identity.

Open

7. OpenCode’s 100T tokens/day are…

What was served on 20 Aug.
A factory capacity sign, not a meter. Observed ~2–4T.
Proof it is ByteDance, because Doubao does 120T.
A typo (it was 100B).

Score. A factory capacity sign, not a meter. Observed ~2–4T.

“We have capacity for 100T/day”. Day 1: ~3.85T combined. OpenRouter ~2T in ~2 days. 86–94% is cache: lookup, not matmul.

Trap. Do not add OpenRouter + OpenCode. Do not ID the lab by the invoice: the CDN can be other than the chef.

Open

8. Why can Ox not be the brochure GLM-5.3?

Because the brochure says 128K out and Ox 131,072 (that is enough).
Because public 5.3 is text-only; Ox sees image and video, and the toll = 5V-Turbo.
Because it has another tokenizer.
Because Z.ai denied it.

Score. Because public 5.3 is text-only; Ox sees image and video, and the toll = 5V-Turbo.

Docs.z.ai: “currently supports text-only inputs”. OpenRouter already lists glm-5.3 at 131,072, so 128K vs 131,072 is not the passport. Modality is: cameras + 5V encoder.

Trap. “Not the brochure one” ≠ “not the family”. Thesis: unpublished V/Flash variant. Z.ai has not spoken.

Open

9. The three data contracts (OpenCode ZDR / sheet “retain, don’t train” / stealth EULA)…

Say the same: zero retention.
Cannot all be true at once. Do not paste secrets.
Only apply if you pay.
Prove the lab is Western (GDPR).

Score. Cannot all be true at once. Do not paste secrets.

OpenCode: ZDR. OR sheet: retained, not trained. Default EULA (6 Jul): collected for training. “This time” is a carve-out on a contract that trains.

Trap. Do not quote OpenCode ZDR as if it were OpenRouter policy. The path matters.

Open

10. Xiaomi MiMo dies as the weights author because…

They never did stealth.
Tokenizer L1=176, video 910≠296, they accept audio, unclecode 2/4 and 0/4.
Manifold says 2% and that is enough.
Z.ai sued them.

Score. Tokenizer L1=176, video 910≠296, they accept audio, unclecode 2/4 and 0/4.

The Hunter/Healer playbook fits. The instruments do not. They already burned that animal. A MiMo-V3 would have to clone the GLM knife + 5V encoder + 1210 + drop audio.

Trap. Modus operandi is not a fingerprint. “They did it before” is a prior, not L1=0.

Open

11. To claim “Ox Alpha IS GLM” at 9.9 you would still need…

Nothing: +75 is already a passport.
Manifold at 99%.
A Z.ai statement or our own live tokenizer/video run. Until then: a strong hypothesis.
Pliny tweeting it.

Score. A Z.ai statement or our own live tokenizer/video run. Until then: a strong hypothesis.

Four kitchens point at the same neighborhood. Nobody at the lab signed. This briefing rebuilds reports; it did not call the API. Inflating identity to 9.9 would drop honesty below 9.9.

Trap. A “I am GLM” jailbreak can be acted. The tripod (+75, 296, 1210) cannot. Still not a press release.

Open

12. “Fewer parameters” in 2026 usually means…

Always dumber.
MoE: a huge building where only some experts wake per token. More efficient, not a toy.
That it is not frontier.
That it fits on a phone.

Score. MoE: a huge building where only some experts wake per token. More efficient, not a toy.

GLM-5: ~744B total, ~40B active. Interconnects: a third of Kimi K3 and still frontier at coding. You pay for who switches on.

Trap. Flash/~300B is rumor, not a sheet. “Fewer” without saying active vs total is a marketing trap.

Open