Volume XI · Reading the Reports
How to Read a Technical Report
Two frontier models were released within a few months of each other. One arrived with a paper giving layer counts, attention mathematics, token budgets and the cost of the training run. The other arrived as a post on a social network. Both were, in the ordinary sense of the word, announced.
The mental model
A model release is a document with a genre, and the genre tells you what the laboratory is willing to be held to. Learn to read the genre before you read the benchmarks.
The gap widened
Through 2025 and into 2026 the disclosure practices of frontier laboratories diverged rather than converged. Some publish architecture, hyperparameters, ablations and training recipes in enough detail that a well-funded group could reproduce the work. Some publish a training corpus and every intermediate checkpoint. Some publish evaluations and safety analysis with essentially nothing reproducible about the model itself. And at least one 2026 frontier release carried no technical report and no model card at all.
This lecture takes no position on whether that is good. It takes the position that the spectrum is a fact about the literature you must read, and that a student who cannot place a document on it will mistake a marketing surface for a technical one. The genres below are the practical taxonomy. Learn to identify the genre in the first minute, because it determines which questions the document can possibly answer.
The full recipe, and the reproducible release
The full recipe is the genre this course has depended on throughout. Architecture, layer counts, the attention mathematics written out, expert counts, data volumes, and often the training economics. DeepSeek-V3 is the model of the form; Kimi K2, GLM-5, MiniMax M2 and M3, Nemotron 3 and Qwen3 all belong to it. Almost everything in Volumes II through VI was learned from documents of this type, which is itself worth noticing: a course on attention mechanisms is only possible because some laboratories publish attention mechanisms.
Rarer, and more valuable to a student than anything else on this list, is the fully reproducible release. Ai2's OLMo 3, published in November 2025, ships the weights, the Dolma 3 training corpus, the training code, every intermediate checkpoint, and tooling for tracing model behaviour back to training data. A full recipe tells you what was done. A reproducible release lets you watch it happen and check. If you want to see what a language model is actually made of, end to end, this is the artefact to open — not because it is the strongest model, but because it is the only one where every question has an answer you can look up rather than infer.
The model card, and what it does not say
The model card documents capabilities, evaluations, intended use and safety, and is silent on how the model is built. Gemini 3.1 Pro's card, Claude's system cards and the GPT-5-series system cards are all of this type. They are careful documents, often more rigorous about evaluation than the architecture papers are. They simply have a different subject.
The teaching exercise here is to notice what a sentence like "a sparse mixture-of-experts model with a one-million-token context window" leaves undetermined. It does not tell you the expert count, the number of experts active per token, or therefore the active parameter count — the number that governs serving cost. It does not tell you the attention mechanism, and after Volume IV you know that is a live architectural argument with at least three answers in production. It does not tell you the positional scheme, so you cannot tell whether the context was trained or extended by the recipe of Lecture 14. It does not tell you the data. Everything this course has taught you to ask about is outside the frame of that sentence.
The forced disclosure, and the announcement
A forced disclosure is what a closed laboratory must reveal in order to ship open weights, because weights on disk are a specification whether or not a paper accompanies them. gpt-oss is the only real window into OpenAI's architectural choices, and it is a detailed one: alternating dense and sliding-window attention layers, learned attention sinks, and expert weights quantised to MXFP4. Read against the closed releases from the same laboratory, it is the nearest thing to evidence about a family of models that otherwise publishes only evaluations.
At the far end sits the announcement. Grok 4.5, released in July 2026, arrived with neither a technical report nor a model card. Stated factually, without inference about motive: for that model, the documentary record available to a reader consists of the announcement and whatever the model itself does under testing. Every architectural question in this course is, for that release, unanswerable from primary sources.
| Genre | Examples | Discloses | Does not |
|---|---|---|---|
| Full recipe | DeepSeek-V3, Kimi K2, GLM-5, MiniMax M2/M3, Nemotron 3, Qwen3 | Architecture, attention, expert counts, data volumes, often training cost | Usually the corpus itself and the checkpoints |
| Fully reproducible | OLMo 3 (Ai2, Nov 2025) | Weights, Dolma 3 corpus, code, every intermediate checkpoint, data attribution | Nothing material — this is the point of the genre |
| Model card | Gemini 3.1 Pro, Claude system cards, GPT-5-series system cards | Capabilities, evaluations, safety analysis, intended use | Expert and active parameter counts, attention, positions, data |
| Forced disclosure | gpt-oss | What open weights necessarily reveal: layer pattern, sinks, quantisation | Training data, training recipe, the closed sibling models |
| Announcement | Grok 4.5 (July 2026) | That the model exists | Everything else |
A checklist to read with
Take a new release and answer these in order. The point of the exercise is not the answers; it is the pattern of blanks left at the end, which is the fastest characterisation of a document's genre available.
Six questions for any release
1. What are the total and the active parameter counts, and therefore the active
fraction?
2. What attention mechanism — and if the model is hybrid, at what ratio of
layers?
3. What is the context length, and was it trained at that length or extended to
it?
4. What numerical precision was used for training, and what precision are the
released weights?
5. What is the post-training recipe, and what was it optimised against?
6. Which of the above does the document simply not say?
Why openness compounds, and how to date your notes
There is a practical argument for open architectural publication that does not require any ideological commitment, and it is visible in this course's own material. DeepSeek published a sparse attention design; GLM-5 adopted it and shipped it in a flagship. One laboratory's paper became another laboratory's architecture within a release cycle. That is a mechanism, not a slogan — and it is why the models discussed in Volume IV are so heavily weighted toward the laboratories that publish. They are the ones whose ideas can be traced.
A caution to close on, since it applies to reading this course as much as to reading a release. For several 2026 models, architectural details circulate through blogs, aggregators and secondhand summaries before any primary paper exists, and some of those details are later corrected or never confirmed. Kimi K3's technical report had not appeared at the time of writing unconfirmed, yet its architecture was already being discussed in detail. Two habits fix this: check whether a claim has a paper behind it before repeating it, and date every note you take. A fact about this field without a date attached is a fact you cannot re-verify, and the next lecture is built entirely out of dated facts.
Read the primary source
- OLMo 3 — Ai2, November 2025.
- gpt-oss — OpenAI, 2025.
- DeepSeek-V3 — 2024.
- Qwen3 — 2025.
- Gemini 3.1 Pro model card — Google DeepMind.