Update: V0 documentation is live

The Lift Wing LLM documentation is now up to date on Wikitech: Machine Learning/LiftWing/Large Language Models.

Documentation structure (team sync, 2026-07-15)

The current page serves a mix of audiences, including external users (community members and volunteer developers) as well as internal users (admins and clients).

We agreed to keep this Wikitech page as the single source of truth containing the complete documentation (both public and internal models). For Wikimania, we'll create a separate program page containing only the information relevant to volunteers/developers, along with a FAQ.

Next steps

• Create the Wikimania program page by extracting the relevant subset of the documentation.
• Expand the Rate Limits & Access section:
• Add guidance for accessing models from Toolforge.
• Document the authentication/sign-in requirements for each platform.
• Confirm the base model and precision details for internal models.

I'm wrapping up the Available Models / Deployment sections of the LLM docs. I pulled precision, tensor-parallel, and context length from the deployment configs, but need to confirm a few base-model identities:

• gpt-oss-safeguard-20b — base = OpenAI gpt-oss-safeguard-20b, ~20B MoE, ~3.6B active, MXFP4?
• cope-b-a4b — what's the actual base model? Gemma-4-26B-A4B MoE, 26B total / 4B active? Precision BF16?
• cope-a-9b (experimental) — Gemma-2-9B LoRA merge, 9B params, BF16?
• aya-expanse-8B — Cohere Aya Expanse 8B, served int4 via bitsandbytes?

Update

New program page for Wikimania is created: LiftWing/Large_Language_Models/Wikimania_2026

It pulls the relevant subset of the main documentation, with links back to it for anyone who wants to dive deeper. Feedback is very welcome! Let me know if you think there's anything else we should add.

As requested on slack, I had a look at the Available Models and Deployments documentation and compared it with the current deployment configs at commit 302df28 then cross-referenced the model cards on Hugging Face. GPUs, precision, and context check out for gpt-oss-safeguard-20b, cope-b-a4b, and cope-a-9b (gpt-oss-safeguard confirmed at the source: 21B total / 3.6B active, MXFP4, fine-tuned from OpenAI's gpt-oss under Apache 2.0), and aya-expanse-8b is correctly absent. Five updates needed:

1. cope-a-9b nodes: the wiki lists ml-serve1012 + 1013, but the config pins only 1013. 1012 was removed by 1319050 (2026-07-29).

2. qwen3-embedding nodes: should be ml-serve1013–1015, not 1012–1015 (same 1319050 change).

3. jina-embedding precision: listed as fp16 in both tables, but it deploys as bf16 (size estimate unchanged).

4. Stale TODO in the page source: the wikitext contains two hidden editor notes (HTML comments, not visible on the rendered page): one below the Available Models table and one below the Deployments table. The first still references aya-expanse-8b, whose service (aya-llm) was removed as unused in 1319050; that reference should be dropped.

5. CoPE base models (resolves the second source TODO): the note below the Deployments table asks for confirmation of the CoPE base-model identities. Confirmed via Hugging Face: cope-a-9b is Zentropi's CoPE-A, built on Google's Gemma-2-9b via LoRA fine-tuning (model card); cope-b-a4b is Zentropi's CoPE-B, a ~25–26B Mixture-of-Experts model using ~4B active parameters per call (Zentropi on HF; also a ROOST Model Community model, like gpt-oss-safeguard). The ~52G bf16 checkpoint recorded in 1308563 supports the 26B total.

All changes suggested in T431847#12182917, have been done. The current documentation now shows:

1. Available Models

2. Deployments

Update — added the last two pieces:

• Main page: added a Data & privacy section — the API persists nothing (prompts and responses aren't logged, retained, or used for training), while LiftWing Studio saves chats in its database by default, with temporary-chat and sharing options.
• Program page: added a collapsible FAQ (models & use cases, capabilities, limits & cost, data & privacy, support).

With that the V0 documentation is complete. Marking as Resolved.