LLM Tech Evolution Matrix

中文版：README.zh.md

A structured, continuously updatable knowledge base that tracks and decomposes the technical stacks of mainstream AI models — for horizontal comparison across vendors and vertical analysis of single techniques over time.

Status

M1 in progress; public accelerator catalog added. 35 LLM/VL records across 5 vendors (DeepSeek, Qwen, Kimi, Z.AI, MiniMax) validate against LLM schema v8. 3 media records (H3, Music3, Speech 2.8) use independent media schema v1; 6 engine snapshots remain separate. Generated technique/coverage and engine-support reports exist; an interpretive trend report is still outstanding.

Newest additions cover DeepSeek-V4.1-Flash-1017, V4-Pro-0813, Flash-Vision-Exp, and MiniMax's Text-01/VL-01 and M1–M3 lineage. Media records distinguish released weights from hosted services. M3.1 Flash is sourced but has no released weights and is not counted as extracted.

Browse the LLM index, media index, and media schema. This refresh's tasks, review and handoff live in the coordination state; the full model inventory is in tasks/ROADMAP.md.

The hardware catalog adds 10 representative accelerators across NVIDIA, AMD, Google, AWS and Intel, using independent hardware schema v1. Public specifications carry source links, exact variants, units and peak-performance qualifications. Internal measurements are excluded. Start with the hardware procedure or active coordination state.

Where to find things

If you want to …	Read
Understand why this project exists	docs/vision.md
Know the LLM/VL extraction fields	docs/schema.md
Understand the sourcing → extraction → synthesis flow	docs/pipeline.md
Follow naming, file layout, schema-versioning rules	docs/conventions.md
See the strategic roadmap (milestones, scope)	docs/roadmap.md
See which models are queued / extracted / reviewed	tasks/ROADMAP.md
Browse LLM/VL records	data/extracted/README.md
See which models use which technique	data/reports/technique-index.md

For media use the schema and index. For engines use the schema, snapshot inventory and support matrix.

Repository layout

docs/                  Authoritative reference docs (vision, schema, pipeline, conventions, roadmap)
  glossary/            Per-technique wiki, bilingual — short entries with "Used by" tables
src/llm_tech_matrix/   Python package
  engine_schema.py     Engine snapshots (schema v3)
  media_schema.py      Media generation (schema v1)
  hardware_schema.py   Public accelerator specifications (schema v1)
  schema.py            Pydantic models — executable version of docs/schema.md
  sourcing/            Layer 1 — fetch HF configs, papers, blogs; manifest + sha256
  extraction/          Layer 2 — agent extraction; render*.py generate bilingual views.
  synthesis/           Layer 3 — registry.py + index.py generate the browse surfaces
scripts/               validate_extractions.py + validate_registry.py (CI gates) + migrations
data/
  sources/<model>/     manifest.json (committed) + cached raw files (gitignored)
  extracted/<model>.json  Schema-validated extraction output (committed)
  extracted/<model>.md    Rendered English summary (committed, generated from the .json)
  extracted/<model>.zh.md Rendered Chinese summary (committed, generated from the .json)
  extracted/engines/     Engine JSON + bilingual summaries (schema v3)
  extracted/media/       Media JSON + bilingual summaries/index (schema v1)
  sources/hardware/      Per-variant public source manifests + ignored cached files
  extracted/hardware/    Hardware JSON + bilingual summaries/index (schema v1)
  extracted/README.md     Generated model index (do not hand-edit)
  reports/             Generated: technique-index.md (technique↔model matrix), coverage.md
tasks/
  projects/            Project/phase coordination state and handoffs
  ENGINES.md           Engine snapshot inventory and calendar
  ROADMAP.md           Per-model status table + Current focus
  models/<model>.md    Per-model notes, sources, open questions
  hardware/<slug>.md   Per-variant source/evidence notes and open questions
.claude/skills/        LLM/engine extraction and PR drafting procedures

Run uv run python -m unittest discover -s tests -v for media and hardware regression checks. CI also runs schema, source-evidence and generated-view validation via scripts/validate_extractions.py.

Quickstart

This project is built with uv on Python 3.13.

# Install the locked dependencies
uv sync --locked

# Validate committed LLM, engine, media and hardware records and generated indexes
uv run python scripts/validate_extractions.py
uv run python scripts/validate_registry.py

Adding a dependency:

uv add <package>

Development setup

Once after cloning:

uv sync                       # install deps (runtime + dev)
uv run pre-commit install     # activate the git hook

Full guide — style, lint/format commands, CI, AI code review, PR conventions — in docs/development.md.

Validating committed records needs no HF_TOKEN. Consult .env.example before fetching sources; model-weight downloads and inference are outside this workflow.

Contributing a new model extraction

1. Add an entry to tasks/ROADMAP.md with status backlog, and create tasks/models/<model-slug>.md listing candidate source URLs.

2. Register the sources — this downloads them, records sha256, and writes the manifest: uv run python -m llm_tech_matrix.sourcing add <slug> --name config --kind hf_config --url ... --filename config.json (then python -m llm_tech_matrix.sourcing.pdf_to_text <slug> if any asset is a PDF).

3. Run the extract-model skill (or extract by hand following docs/schema.md) to produce data/extracted/<model-slug>.json.

4. Validate it against src/llm_tech_matrix/schema.py — this is what CI enforces.

5. Render the committed bilingual summaries: uv run python -m llm_tech_matrix.extraction.render <slug>. Never hand-edit the generated .md / .zh.md — change the JSON or the renderer instead.

6. Add "Used by" rows to the relevant docs/glossary/ entries (both <entry>.md and <entry>.zh.md), and update the roadmap status to extracted.

7. After vocabulary/glossary updates, run uv run python -m llm_tech_matrix.synthesis.index, then both validation commands above so model indexes and technique reports stay current.

These steps are for LLM/VL records. Audio/video generation uses the media procedure and data/extracted/media/; engine snapshots use the engine procedure. Accelerator specifications use the hardware procedure, --track hardware sourcing and data/extracted/hardware/.

The full LLM/VL procedure lives in .claude/skills/extract-model/SKILL.md.

Cardinal rule

When information is not present in source material, the field value is the literal string "[Unknown/Not Disclosed]". Never hallucinate. Half this project's value is data you can trust. See docs/schema.md for the full rule and the inferred-fields mechanism for closed models.