--- license: apache-2.0 library_name: gguf tags: - gguf - llama.cpp - local - vision - audio - offline --- # LYGO Console Models — the weights behind the LYGO LLM Console (V1) Weights store for the offline, local-first **LYGO LLM Console** (PC LOCAL and USB LOCAL) and for the light installers, which fetch from **this repo only** — never from a third-party mirror, never from a moving branch. All builds and installers are listed at ****. Model host: this repo · GitHub mirror: ## Files | file | bytes | SHA-256 | licence / origin | |---|---|---|---| | `gemma4-12b.gguf` | 7,381,382,048 | `1278394b693672ac2799eadc9a83fd98259a6a88a40acfb1dcaa6c6fc895a606` | Gemma 4 12B Unified, Apache-2.0 (Google) — text, image, audio | | `gemma4-12b-mmproj.gguf` | 175,115,584 | `675ad6e68101ca9413ec806855c452362f0213f2dfc5800996b086fdb8119842` | vision/audio projector for the above | | `qwen2.5-coder-7b.gguf` | 4,683,074,048 | `60e05f21…` (see `models.lock.json`) | Qwen2.5-Coder-7B-Instruct, Apache-2.0 (Alibaba) | | `nomic-embed-text-latest.gguf` | 274,290,656 | `970aa74c…` (see `models.lock.json`) | nomic-embed-text, Apache-2.0 | | `models.lock.json` | 3,883 | — | pinned revisions + digests for the fetcher | | `fetch_models.py` | 11,248 | — | the fetcher the light installers ship | | `LICENSE-APACHE-2.0.txt` | 11,358 | — | Apache License 2.0 (full text) | `models.lock.json` is the authority: it pins immutable revisions (`…/resolve//…`) and a SHA-256 per file. The digests above are what the console verifies at boot. ## Fetch, don't click ``` python fetch_models.py --list python fetch_models.py --profile basic --dest D:\LYGO_MODELS # gemma4-12b + projector python fetch_models.py --profile full --dest D:\LYGO_MODELS # all four ``` The fetcher resolves every URL through the pinned revision in `models.lock.json`, downloads from this repo, verifies SHA-256, and only then keeps the file. `--check --dest ` verifies an existing collection without downloading. ## Sizing The console's rule is `model size × 1.6 + 2 GiB ≤ max(free RAM, half of installed)`. gemma4-12b needs ~32 GB RAM (or a GPU with VRAM to spare — ~80 of 99 layers fit an 8 GB card). On 8–16 GB machines use a 1.5B–3B Q4 model instead; on 64 GB+ go bigger. The console boots whatever you put in `models\`, so a better system is a drop-in upgrade with no reinstall. ## Licences and attribution Model weights are Apache-2.0 as published by their authors; the full Apache-2.0 text is included here. **Gemma** is a trademark of Google LLC; **Qwen** of Alibaba Group. These weights are redistributed unmodified with attribution. This repo is **not affiliated with or endorsed by** Google, Alibaba, Meta, or the llama.cpp project (engine: llama.cpp b10988, MIT). The console source is licensed under the **LYGO Sovereign License v3.0** — free to use and build on; not MIT; not for resale, rebranding or white-labelling. Legal summary: Nothing here is medical, legal or financial advice. AI output can be confidently wrong — verify before you act on it. Steward: **Justin Helmer** (Lightfather / Excavationpro) · signature `Δ9Φ963`