Searching Models
hali gives you two search surfaces: the CLI (against HuggingFace directly) and the public registry at hali.network.
CLI search
Search HuggingFace for GGUF models from your terminal:
hali search mistral
hali search "llama 3 instruct"
hali search codellama
Results are ranked by download count, filtered to GGUF-only repos:
1 TheBloke/Mistral-7B-Instruct-v0.2-GGUF 2.1M downloads
2 bartowski/Mistral-7B-Instruct-v0.3-GGUF 840.5K downloads
3 MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF 320.1K downloads
4 lmstudio-community/Mistral-7B-Instruct-v0.3-GGUF 210K downloads
...
Use 'hali pull <repo>' to download a model.
How CLI search works
- Queries HuggingFace's API directly — no hali infrastructure involved
- Filters to GGUF format only
- Sorts by download count (most popular first)
- Shows the top 20 results
From search to pull
After seeing results, you can:
# Pull interactively — hali re-runs the search and lets you pick
hali pull mistral
# Pull a specific repo directly
hali pull TheBloke/Mistral-7B-Instruct-v0.2-GGUF
# Pull by canonical ID — if cached, instant. If not, hali searches HF automatically
hali pull mistral:7b:instruct:q4_k_m
Registry search (hali.network)
The public registry at hali.network searches models that have been submitted as signed manifests by publishers. This is a curated catalog — not everything on HuggingFace is in the registry, but everything in the registry is verified.
What the registry shows
- Model name and format (GGUF, safetensors, etc.)
- Trust score (0.0–1.0) — computed from format validity, publisher reputation, download success rate, and age
- Publisher identity — 64-char Ed25519 public key with optional display name
- Infohash — the BitTorrent identity (v1 40-char SHA1, v2 64-char SHA256)
- Files — list of files included in the model
- Magnet URI — copy-pasteable link for any BitTorrent client
Registry vs. HuggingFace
| CL I search | Registry search | |
|---|---|---|
| Source | HuggingFace API | hali-backend PostgreSQL |
| Scope | Every GGUF on HF | Models published via hali |
| Results | Download-ranked | Trust-score-ranked |
| Verification | HF metadata | Ed25519 signed manifests |
| Magnet URI | No | Yes |
| Offline | No (needs HF API) | No (needs registry API) |
| Use for | Finding models to pull | Finding verified torrents to join |
The registry trust signal
Models in the registry carry a trust score, which serves as a quality signal:
| Score | Tier | What it means |
|---|---|---|
| > 0.85 | Featured | Verified format, trusted publisher, high download success rate, mature |
| 0.60–0.85 | Normal | Standard listing — meets all basic criteria |
| 0.30–0.60 | Low Rank | New or unproven — score will improve as peers verify it |
| < 0.30 | Hidden | Doesn't appear in search results |
Scores are recomputed every 5 minutes by the backend scoring worker.
Advanced: Finding LAN-available models
If the daemon is running, check what models are available on your LAN without touching the internet:
hali network seen
Shows recent LAN announcements — models currently being seeded by other machines on your local network:
MODEL ID NODE AGE
--------------------------------- -------- ----
mistral:7b:instruct:q4_k_m a3f29b... 12s
llama:8b:instruct:q5_k_m f8c41d... 45s
Use the -p flag to pick one interactively and download it straight from the LAN peer:
hali network seen -p
Or pull directly from LAN (no HuggingFace fallback):
hali network pull mistral:7b:instruct:q4_k_m
If no LAN peer has it, network pull fails with an error — it won't fall back to the internet.
Tips
- Use
hali searchto discover models, thenhali pullwith the repo or model ID to download - Browse hali.network for verified models with trust scores and magnet links
- Use
hali network seento see what your teammates are already seeding — you might not need the internet at all - Canonical model IDs (
mistral:7b:instruct:q4_k_m) are the fastest way to pull, especially in scripts