Local models

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

Sanad model (sole tier)

Model ID	Checkpoint	Tier	VRAM (approx.)	Hugging Face
`nassila-sanad-9b`	FT-6	Sole tier (Qwen 3.5) — pick your quant	Q2_K ~3.6 GB → Q8_0 ~8.9 GB (+ MTP variants)	nassila-sanad-9b

Six default GGUFs (Q2_K / Q3_K_M / Q4_K_M / Q5_K_M / Q6_K / Q8_0) share the same FT-6 trunk (converted with --no-mtp). Matching MTP files (nassila-sanad-9b-mtp-*.gguf) are for llama-server speculative decode. Typical filenames: nassila-sanad-9b-q4_k_m.gguf, nassila-sanad-9b-q6_k.gguf.

Retired (abstract-era): nassila-sanad-4b (S15), nassila-sanad-12b (S14), nassila-sanad-e4b (S12) — their cards stay on Hugging Face for legacy downloads only.

Tier behavior in the app

Tier chip	Checkpoint	When to use
9B	FT-6	Sole Sanad tier — you pick the quant

Switch 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 — full-text source content (registry abstract, OA full text, or attached source PDF)
Structured JSON output — claims, verdicts, verbatim quotes when supported

Sanad grounds against full-text source excerpts. Masdar 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 the Sanad 9B GGUF above.