Local models

Nassila ships Sanad passage grounding against local GGUF models you download yourself. Multi-GB weights are not bundled in the installer.

Sanad checkpoints (S = Sanad)

Checkpoint	Model ID	Tier	VRAM (approx.)	Hugging Face
S12	`nassila-sanad-e4b`	E4B default	~8 GB	nassila-sanad-e4b
S14	`nassila-sanad-12b`	12B quality	~12 GB+	nassila-sanad-12b

Typical GGUF filenames: nassila-sanad-e4b-q6_k.gguf, nassila-sanad-12b-q6_k.gguf.

S12 and S14 label separate ship artifacts (default E4B vs quality 12B) — not a single linear counter on one file.

Tier behavior in the app

Tier chip	Checkpoint	When to use
E4B	S12	Default; laptop-friendly; recommended starting point
12B	S14	Higher-quality grounding when you have sufficient VRAM/RAM

Switch tiers from the Sanad bar tier chip or Settings → Passage grounding.

Runners

Sanad talks to any OpenAI-compatible /v1/chat/completions endpoint:

Runner	Default base URL
LM Studio	`http://localhost:1234`
Ollama	`http://localhost:11434`
vLLM	Your server URL
Custom	User-defined

Setup walkthrough: Sanad setup.

What Sanad receives at inference

Your passage — text around the in-text citation (capped by engine limits)
Source excerpt — registry abstract or OA chunk (up to engine excerpt cap)
Structured JSON output — claims, verdicts, verbatim quotes when supported

Excerpt type in current product builds is primarily abstract-first; OA full text is used when manuscript source fetch succeeds. Masdar-lite (app 1.2.0, shipped) adds open-access PDF text when Unpaywall finds a PDF — no new GGUF required.

Distribution policy

Bring your own GGUF or use Ollama pull from Hugging Face Hub
No automatic download of model weights without your action
Cloud API preset optional — you supply endpoint and credentials

Future

A merged multi-worker bundle may ship later. Today, production grounding uses Sanad-only S12 / S14 GGUFs above.