Formats and routes
GGUF, Safetensors, and MLX bundles across Hugging Face, HTTPS, Iroh, and BitTorrent.
Noema Atlas v0.2.0 · Released July 11, 2026
Find, download, verify, and share local AI models.
Noema Atlas pulls GGUF, Safetensors, and MLX weights from Hugging Face, HTTPS mirrors, Iroh peers, and BitTorrent swarms—then verifies the complete file before it enters your library.
Free · Apache 2.0 · macOS, Windows, and Linux
Current networkRegistry live
See the whole route, not just a progress bar.
Discover a model, inspect its sources, control the queue, and decide when a verified local copy should be shared.
Find the right model.
Formats, quantizations, licenses, and available routes stay visible before a download begins.
Qwen3 8B · Q4_K_MGGUF · Apache 2.0 · 5.2 GB
HFHTTPSIroh · 3BT · 8
Inspect ↗Mistral Small 3.2Safetensors · Apache 2.0 · 48.4 GB
HFIroh · 1BT · 4
Inspect ↗DeepSeek R1 Distill 7BMLX · MIT · 4.1 GB
HFHTTPS
Inspect ↗One transfer. Every eligible route.
Pause, resume, retry, stop, or reorder without giving up verified progress.
Qwen3-8B-Q4_K_M.gguf3.54 GB of 5.20 GB · 68 MB/s · 00:24 remaining
68%Hugging Face32 MB/s
HTTPS mirrorStandby
Iroh peers21 MB/s · 3
BitTorrent15 MB/s · 8
PauseStopMove to topPeers & routes
Keep one verified copy.
Import local files, create share links, or hand a model to LM Studio and Ollama without another download.
Qwen3 8B · Q4_K_MVerified · Shared worldwide · Ratio 1.34
Share on12 peers
Open ↗Llama 3.2 3B InstructVerified · Private import · Stored once
LM StudioOllama
Reveal ↗Flux.1 Dev · MLX bundleVerified · 3 files · 0 duplicate bytes
Local only
Share ↗Illustrative product tour based on the shipping v0.2.0 controls and route data.
Four routes, one result
Every source contributes bytes. None defines the truth.
Hugging Face, HTTPS mirrors, Iroh peers, and BitTorrent swarms can all contribute bytes to the same verified file.
Hugging FaceSearch the Hub, pin revisions, and use authenticated access where required.
HTTPSResume from publisher mirrors with parallel range requests and automatic failover.
IrohConnect directly to peers over QUIC with NAT traversal and multi-peer aggregation.
BitTorrentUse DHT, TCP or µTP, magnets, seeding, peer visibility, and explicit bandwidth controls.
An early public network, measured honestly.
These counters report the live public registry—not private imports, direct downloads, or the models already stored on people’s machines. Atlas is useful as a downloader before the public mesh grows.
Files shared—
Peers online—
Signed manifests—
content-idsha256 → the same verified file, whoever serves it
What shipped in v0.2.0.
The July release turns Atlas into a complete model-distribution workspace: more formats, more routes, and direct control over every transfer.
Transfer control
Pause, resume, stop, retry, and reorder the queue while retaining resumable progress.
BitTorrent, visible
Inspect live peers and magnets; set upload ratios, sequential mode, queue priority, and scheduled bandwidth limits.
Import and handoff
Drop in local files, create share links, then open verified models in LM Studio or Ollama.
Less duplicated data
Content-addressed storage keeps identical bytes once; reflink and hardlink installs avoid extra copies. Signed updates keep the apps current.
Trust and privacy
Know what is verified—and what is visible.
Atlas verifies the complete file before it enters your library. Mismatched content is quarantined and the source is rejected.
- BitTorrent
- Your IP is visible to swarm participants
- Public trackers
- Optional and user-controlled
- Private / gated imports
- Not automatically broadcast
- Commit gate
- Full-file BLAKE3 + SHA-256
Two interfaces. One engine and model library.
Noema Atlas and Noema Atlas Studio share the core, content store, and installed models. Each app keeps its own preferences.
| Compare | Noema Atlas | Noema Atlas Studio |
|---|---|---|
| Interface | Native desktop UI | System-webview UI; no bundled Chromium |
| Footprint | Lightest option | More visual workspace with a modestly heavier footprint |
| Platforms | macOS, Windows, Linux | macOS, Windows, Linux |
| Dependencies | No browser runtime | WebKit on macOS, WebView2 on Windows, webkit2gtk on Linux |
| Shared | Verified engine, content-addressed store, installed models, Iroh and BitTorrent networking | |
| Separate | Native app preferences | Studio preferences in ~/.noema/studio-settings.json |
Noema Atlas
The lean native desktop app. No browser and no Electron underneath it—just the complete network in a lighter footprint.
Noema Atlas Studio
The same engine behind a more visual system-webview interface for people who prefer a richer library and download workspace.
v0.2.0 · July 11, 2026 · Read the release notes · View every release asset
Publish models people can verify.
01 noema keygen
02 noema manifest build --name "My Model" --artifact "model.gguf=/path/model.gguf" \
--source "model.gguf:hf:Org/Repo@main/model.gguf" --sign <key_id> --out model.json
03 curl --fail -H 'content-type: application/json' --data-binary @model.json \
https://atlas.noemaai.com/manifestsA signed manifest binds model identity, license, hashes, and eligible sources. The registry publishes the claim; every Atlas client verifies it locally.
Read the manifest spec