Skip to content

Model Deep Dive guide

Research, explain, illustrate, and review a model study from reports and released artifacts.

Updated View as Markdown

A Model Deep Dive explains what a model changes, how it works, and what the evidence supports. Help readers develop a mental model, not merely read a shorter technical report.

Begin with the defining ideas and useful scale numbers. Connect them to a clear architecture view, then explain the important mechanisms. Use original report figures where they help. Review the technical content and rendered presentation before considering the study complete.

This guide defines the model study workflow for LLM & VLM. Use it with the page template and Content standards. The shared standards govern citations, equations, figure numbering, reuse rights, and public sharing.

The study workflow

Step Required result
Scope A named generation, representative checkpoint, learning goal, and verification boundary
Inspect The complete report, relevant appendices, released artifacts, and targeted implementation
Trace An explanation of inputs, computation, training stages, inference state, and evidence
Illustrate Purposeful source figures or independent diagrams with clear provenance
Review A coverage audit, claim checks, and a refined reading experience
Validate Working navigation, readable layouts, complete exports, and recorded checks

These are Latent Light’s authoring standards. They are not requirements imposed by model publishers.

Select and scope the model

Prioritize recognized LLM and VLM generations with public weights and substantive technical evidence. Choose a model that establishes a baseline, introduces a useful mechanism, or provides a meaningful contrast. Recognition is an editorial judgment, not a fact established by a launch announcement.

Use one page per generation. Compare related sizes and training variants within that page when they share the central lesson. Create a separate study only when the differences justify one. Do not create a thin page for every checkpoint.

Identify the kind of report

Record what the linked document actually covers:

  • Full technical report: architecture, training, and evaluation of the studied generation.
  • Substantive model card: sufficient technical disclosure, despite a different document title.
  • Mechanism or backbone paper: a specific component or experiment, not necessarily the complete released model.

A partial report can support a useful study. State its boundary near the opening and throughout the relevant claims. Do not assign a small experimental model’s recipe or results to a larger released checkpoint. Identify details supplied by the released configuration or implementation separately.

Record documentation-only candidates in docs/research/architectures/coverage.md when the evidence does not support a complete study. State which sources were checked and which evidence remains missing. Do not claim that no report exists after a limited search. Do not publish an empty placeholder page.

Models available only through APIs or restricted partner access can remain in the evolution map. A public report alone does not make them part of this released-artifact collection.

Describe openness precisely

Check weights, inference code, training code, data information, and licenses separately. Do not assume that a public repository makes the complete system open source. OSI distinguishes software distribution rights from the data information, code, and parameters required by its AI definition.12

Use public weights or open weights under the stated license when that is what you verified. If classifying a system as open source, identify the definition and check the required artifacts. Separate access requirements, usage restrictions, and training disclosure. A checkpoint license does not automatically license paper figures.

Define the study contract

Before drafting, record the exact subject and verification depth.

Record Identify
Model Generation, representative checkpoint, size, and base or post-trained variant
Report Title, document type, canonical URL, edition, and publication date
Artifacts Weights, configuration, tokenizer or processor, and separate licenses
Implementation Repository revision, library version, and inspected functions
Scope Modalities, mechanisms, exclusions, and any report-to-release mismatch
Verification Sources read, artifacts inspected, code traced, and experiments actually executed
Dates Release event, source inspection, page revision, and review dates

Use immutable source revisions where possible. Hugging Face supports downloads at a specific revision.3 Inspect model cards for intended use and disclosed limitations, then check their consequential claims against the relevant evidence.4

Keep hashes, checksums, and extraction coordinates in research records and immutable link targets. Do not display opaque hashes in article text, resource cards, tables, or reference labels. Use readable labels such as report v2, the original release, or a named library version.

State the verification boundary near the opening. For example: “Report and configuration inspected. Targeted cache code traced. No checkpoint executed.” Do not use a general verified badge when only some layers were checked.

Keep evidence layers separate

Evidence Can support Does not automatically establish
Family report Disclosed family methods and recipes Every size or later release uses the same settings
Released artifact Its configuration and interface Its complete training history
Controlled ablation The tested change under stated controls The gain on every released checkpoint
Inspected implementation The traced code path in that version Successful execution or optimized backend behavior
Runtime measurement Performance under recorded conditions Performance on other hardware, loads, or lengths
Our derivation or interpretation A calculation or explanatory hypothesis A published finding or measured result

Record conflicts instead of silently choosing a convenient value. Explain the discrepancy when it affects the reader’s conclusion.

Build a useful opening

Place ModelResources before the opening paragraph. Link the original technical document and representative Hugging Face repository directly. Name a model card or mechanism paper accurately in the study scope.

Explain the task and central idea before introducing acronyms or tensor notation. Give two to four highlights: what differs, which problem it addresses, and why the model matters. Support claims of adoption or influence. Label our interpretation of potential value separately.

Keep the opening useful in about a minute. Do not replace it with a launch feature list or a configuration dump.

Model in numbers

Name the checkpoint before its numbers. Keep a compact table with values, units, scope, and nearby citations.

Dimension Important distinctions
Parameters Rounded model name, stored total, active parameters, and counted components
Pretraining data Unique examples, processed tokens, repeated exposure, and modalities
Post-training data Examples, preference pairs, and rollouts by stage
Compute Hardware count, duration, accelerator hours, FLOPs, and covered stages
Context Training length, configuration ceiling, accepted input, generation allowance, and evaluated length
Inference footprint Precision, backend, hardware, batch, load, input length, and output length

Define “active” when reporting sparse model parameters. Identify whether embeddings, output projections, or other components are counted. Do not add stage token budgets unless their definitions make them additive. Hardware count without duration does not establish training compute or financial cost. Do not present a stage-specific rental estimate as the complete training cost.

Mark important missing information undisclosed, not zero. State assumptions beside local derivations. Use a simple two-column table when it improves mobile reading. Move secondary configuration details into the technical sections.

Explain the model, not just its feature names

Use the following structure flexibly. Combine or rename sections when that improves the explanation.

Section Reader outcome
What changed Understand the motivating problem and appropriate predecessor
Architecture Follow representative inputs through the network to its outputs
Distinctive mechanisms Understand computation, costs, and evidence for the defining choices
Training Follow disclosed objectives, data, updated modules, and stage transitions
Inference Understand preparation, state, decoding, and the external serving boundary
Evaluation Interpret comparisons and ablations under their actual conditions
Limitations and lessons Recognize unknowns and transfer useful ideas without overstating them

For each central mechanism, explain its motivation, intuition, computation, tradeoffs, and supporting evidence. Define shapes or notation where useful. Separate the authors’ motivation from our interpretation. Link shared theory to Foundations or Architectural Evolution. Do not repeat a generic Transformer tutorial on every model page.

Architecture and retained state

Trace a representative input through preparation, embeddings, blocks, and output heads. Identify the network, tokenizer or processor, generation wrapper, and serving system separately.

Technical coverage checklist
  • Tokenizer, vocabulary, special tokens, embeddings, output heads, and weight tying.
  • Block order, residual branches, normalization, position representation, attention, and FFNs.
  • Query and KV dimensions, projection widths, masks, and output shape.
  • Expert selection, routing scores, normalization, shared experts, gates, biases, and auxiliary losses.
  • Recurrent state, retained KV, sparse reads, windows, compression, and discarded history.
  • Parameter counts and the components included in total or active computation.

Explain which cost a mechanism changes. Sparse reads do not necessarily remove retained history or every quadratic operation. Do not confuse an activation’s internal skip term with an outer residual branch. Distinguish theoretical operation counts from measured latency or memory.

Training and post-training

Explain the actual disclosed stages, not a plausible generic recipe. For each stage, identify the starting checkpoint, objective, data, updated modules, optimization settings, and output.

Training coverage checklist
  • Data sources, mixtures, filtering, deduplication, synthetic data, and contamination controls.
  • Token budgets, context curriculum, batch definition, optimizer, learning rate, precision, parallelism, and stability measures.
  • SFT targets, loss masks, chat format, response lengths, and token or example weighting.
  • Preference or RL data, rollouts, rewards, reference policy, regularization, and rejection rules.
  • Distillation sources and targets, frozen modules, and transitions between stages.

When an objective is central, inspect its actual computation. Check detached weights, clipping, normalization, sampling, curriculum, and selection rules where the source discloses them. Distinguish a reward signal from the optimization objective. Do not infer PPO, GRPO, or online rollouts from the label “reasoning model.”

Identify human, model, rule, or executable reward sources when disclosed. Record selection or judge limitations when they affect the evidence. Separate a reported recipe from a proposed reproduction. Mark material missing settings as undisclosed.

Inference and interfaces

Trace input preparation, prefill, retained state, decoding, and output handling. Use the appropriate stages for nonautoregressive models.

Inference coverage checklist
  • Processor settings, chat templates, role tokens, positions, and masks.
  • State or cache inputs and outputs, history growth, eviction, and recurrent updates.
  • Sampling, reasoning modes, structured output, stopping, and tool handoff delimiters.
  • Precision, quantization, kernels, batching, parallelism, and supported backends.
  • Hardware requirements, supported modalities, and input and generation limits.

Tool calls require an external execution harness. Do not present that harness as part of the network. Match reasoning and chat behavior to the documented artifact and source version. Separate reference code from optimized serving implementations.

For performance results, retain hardware, software versions, precision, batch, concurrency, and sequence lengths. Distinguish first-token latency, decode rate, end-to-end latency, and throughput. A kernel timing is not a complete model measurement.

Label examples as conceptual, code-inspected, or executed. A toy implementation is not a checkpoint reproduction.

Additional coverage for VLMs

Explain how visual evidence enters the language computation, not merely that a vision encoder exists.

Multimodal coverage checklist
  • Resizing, crops, tiles, resolution settings, patching, and a defined token-count example.
  • Vision encoder, feature levels, merger or projector, and feature injection sites.
  • Image, text, and video ordering, position coordinates, masks, and special tokens.
  • Frame selection, timestamps, temporal sampling, and temporal representation.
  • Alignment stages, multimodal mixtures, frozen modules, objectives, and loss assignment.
  • Relevant evaluation and failures in OCR, documents, charts, spatial reasoning, or video.

Distinguish pixels, encoder patches, merged features, wrapper tokens, and final sequence positions. State processor settings and which elements a token count includes. Do not assume that feature injection adds sequence positions. Separate default configuration values from settings used in a particular example or evaluation.

Audit coverage and evaluation

Read the complete report and relevant appendices before declaring coverage complete. Maintain coverage-audit.md in the study’s research directory. Map important mechanisms to source locators, page explanations, and any justified omissions. Include preprocessing, positions, objectives, curriculum, inference state, ablations, and failure evidence where relevant.

Inventory useful report figures separately. Record each figure’s purpose, selected asset, or reason for omission. A broad overview does not replace a missing explanation of routing, visual geometry, or training feedback.

Select evaluations that test the central claims. Preserve the checkpoint, benchmark version, metric, baseline, conditions, and source locator. Retain prompting, sampling, attempts, reasoning budget, tools, and judge when they affect comparisons. Distinguish absolute scores, relative changes, and percentage-point gains.

Label results author-reported, externally reported, or measured here. Code inspection is not independent replication. For ablations, identify what changed, what remained fixed, and which model was tested. Do not assign an internal ablation gain to every released model. Explain confounds and missing evaluation details. Seek failures and counterevidence, not only supporting results.

Use original figures deliberately

Inspect the original architecture figure first. Prefer it when it teaches the studied generation clearly and has a documented reuse basis. Use additional mechanism, training, or evaluation figures only when they teach distinct lessons. There is no required figure count.

Follow the paper figure reuse standard. Inspect the document’s terms, permission, or an applicable exception. Missing Creative Commons terms alone do not require omission. If no defensible basis is established, link the original and create an independent mechanism view. Record the reason. Do not replace a useful source figure with generic flow cards for convenience.

Extract, attribute, and render

  • Preserve the exact report edition and checksum. Do not change the archived PDF.
  • Extract a high-resolution figure. Crop surrounding whitespace or prose, not labels, arrows, legends, axes, or essential panels.
  • Record the page, original figure number, extraction method, crop, and reuse basis.
  • Save reviewed derivatives under public/figures/model-studies/<model-slug>/.
  • Keep public-safe asset records in SOURCES.md and applicable notices in NOTICE.md.
  • Attribute original artwork beside the figure. Preserve required copyright and license notices.

Use FocusCanvas with SourceFigure for report artwork. Provide an image slot, meaningful alternative text, and a concise numbered caption. Keep the paper’s figure number separate from our sequential page numbering. Explain what the reader should notice beside the visual, not in an overloaded caption.

Set minWidth for dense figures after checking readable labels. Show the complete figure on desktop. On smaller screens, contain panning within the figure frame. Check keyboard scrolling and the full-size link. Do not let the image widen the entire page. A full-size link does not excuse unreadable inline artwork.

Check light, dark, and focus views. Preserve scientific colors and relationships unless documenting an adaptation. Original white-background report artwork can remain white in dark mode. Review embedded photographs or other third-party elements separately. Permission to reproduce a figure does not authorize publishing the complete report.

Fill a genuine visual gap

Create an independent SVG, HTML diagram, or ModelFlow only when it clarifies a missing or unsuitable source view. Show execution order, shapes, branches, feature paths, or state boundaries where they prevent confusion. Label the visual as our synthesis and cite the sources behind it. Use generated illustrations for intuition, not as evidence for tensor contracts or benchmark values. Do not add decorative visuals to meet a visual quota.

Use small code excerpts only when they resolve a technical question. Identify the source version and function, explain the relevant lines, and retain required notices. Link instead when copying provides no additional learning value.

Keep research records complete

The published MDX page is each new study’s canonical explanation. Keep its evidence under docs/research/architectures/<model-slug>/, with a coverage-audit.md and source, claim, run, and visual records. Use state/task_spec.md and state/progress.json to record scope, the canonical page path, and actual progress. Follow the Research workflow and its repo-local skill’s schema 3 output contract. Do not maintain a duplicate narrative report.

A requested standalone report can use report.md as its canonical output. Existing schema 2 studies retain their historical report-based contract. Preserve those reports and records; do not bulk-migrate them or replace them with empty placeholders.

Record exact source revisions, inspected functions, claim locators, conflicts, asset provenance, commands, and real review results. Keep every research file safe for a public repository. Internal means outside the site, not private. Keep restricted or unclear-rights PDFs outside the public repository. Do not commit weights, datasets, secrets, or unnecessary large media.

Mark a study complete or reviewed only after the recorded checks actually occurred. Separate source-inspection dates from later layout or export checks. A collection-wide regression check does not mean every technical claim was freshly rechecked. Record unavailable reviews as gaps, not successful checks.

Runtime experiments are optional. Do not download large weights, execute untrusted code, provision compute, or spend money without appropriate authorization. If testing is requested, record the environment, commands, outputs, and measurement conditions.

Review and refine before handoff

Review technical claims and presentation in separate passes. Then read the page as one explanation.

  • Can a new reader explain the distinctive idea after reading the opening?
  • Are the model numbers useful, correctly scoped, and easy to compare?
  • Do the mechanisms connect naturally to training, inference, and evidence?
  • Does each figure teach something, with readable labels and a short caption?
  • Can repetition, jargon, clutter, or secondary detail be removed without losing meaning?
  • Does the explanation remain complete on mobile and in text exports?

Fix the identified issues and inspect the result again. Repeat after material revisions. Do not stop at the first draft or a passing build. Consider the study ready only when its explanation is accurate and its presentation is worth sharing.

Completion checklist

  1. Opening resource links, highlights, relevance, checkpoint, and verification scope are clear.
  2. Numbers retain units, stages, parameter definitions, context boundaries, and nearby citations.
  3. Important mechanisms, training disclosures, inference behavior, and unknowns are covered.
  4. Findings, implementation facts, derivations, interpretations, and measurements remain distinct.
  5. Evaluation claims preserve controls and do not overstate causality.
  6. Figures and equations have sequential numbers, stable anchors, accurate provenance, and readable rendering.
  7. Sources and assets satisfy the shared public-sharing standards.
  8. The coverage audit and canonical research records match the completed work.
  9. Desktop and mobile views, both themes, focus, and relevant keyboard behavior were inspected.
  10. Navigation, search, Markdown, MDX, /llms.txt, and /llms-full.txt include the intended content.
  11. Automated validation passes. Remaining limitations and pre-existing warnings are reported.

Use the Maintenance checks and the template’s rendering recipes. Check actual equation parsing, not only numbering patterns. Check that exports retain figure links, captions, references, equations, and diagram text equivalents. Raw component markup or diagram JSON must not replace the explanation in Markdown.

Expand and maintain the collection

Group related generations by family in Model Deep Dives. Keep single-study entries directly in that group. Put thematic reading paths on the overview page. A model can belong to several paths without duplicated sidebar entries. Preserve established model URLs when reorganizing navigation. Keep authoring guides and templates in the collapsed Content Guidance group within About.

Complete and review one pilot study before expanding a batch. Maintain the coverage register with candidates, report scope, artifacts, status, and the next missing check. Choose additions for learning value, not recency alone.

Use parallel agents only when the user authorizes delegation. Assign separate model or family files, assets, and research directories. Keep shared navigation, components, tests, and integration under one owner’s control. Use a separate review pass. Preserve other contributors’ changes and report unresolved work accurately.

For updates, identify the changed source or artifact. Revise the affected explanation, figures, numbers, and comparisons together. Update the page date after a completed material revision. Preserve older evidence dates when that evidence was not rechecked. Do not schedule refreshes, commit, push, or deploy without authorization.

References

  1. Open Source Initiative. The Open Source Definition, version 1.9. Inspected October 10, 2026. ↩

  2. Open Source Initiative. The Open Source AI Definition, version 1.0. Inspected October 10, 2026. ↩

  3. Hugging Face. Download files from the Hub, “From specific version.” Inspected October 10, 2026. ↩

  4. Hugging Face. Model Cards, intended use, limitations, and licensing metadata. Inspected October 10, 2026. ↩

Navigation

Type to search…

↑↓ navigate↵ selectEsc close