Appendix

Model by Model

Every lecture in this series draws its numbers from somewhere. This appendix is that somewhere, collected into one table per laboratory: release date, parameter count, attention mechanism, context length, training precision, and — the two columns most reports leave out — what precision the weights were actually released in, and what a server needs to run them. A blank cell is not an omission on this page's part; it is the fact. The laboratory did not say. Lecture 23 argues that the shape of the gaps is itself the most honest thing a report tells you — this table is where that argument becomes checkable.

---

DeepSeek

Qwen — Alibaba

Kimi — Moonshot AI

Meta — Llama

Google — Gemini & Gemma

OpenAI

Anthropic

Zero architecture rows in this table have a single disclosed cell — that emptiness, across an entire laboratory, is itself the data point Lecture 23 asks you to notice.

Mistral

GLM — Zhipu / Z.ai

MiniMax

xAI — Grok

NVIDIA — Nemotron

Ai2 — OLMo

Other notable releases

---

What the blanks add up to

Count the disclosed "Released / inference" cells above: four labs state a released or serving precision in a primary source — OpenAI (MXFP4, gpt-oss only, its one forced-open release), NVIDIA (NVFP4, the whole Nemotron 3 line), Kimi (INT4 QAT, the Thinking variant only), and Ai2 (native, because OLMo 3 ships everything). Every frontier closed lab — OpenAI's flagship line, Anthropic, Google's Gemini, xAI — discloses nothing about the number format its weights actually run in. That is not this page failing to find the fact. It is the fact.