Release notes and security advisories. Nothing else.

Open source, and helpers on Windows, macOS and Linux

MFabric Consumer is now open source under the GNU AGPL v3.0, with a commercial license available. The source archive is on the download page and inside the Windows installer.

• Use a model on your strongest computer from any other computer you own: a Windows, macOS or Linux helper serves it after you approve it on the host.
• mfab helper install-llama-server installs the official llama.cpp build for the computer (CUDA, Metal or CPU) and checks every file's published SHA-256.
• Approvals and helper connections survive a host restart; revoking a helper refuses it immediately.
• The host checks it can reach a helper and prints the exact fix for Windows Firewall or macOS Local Network permission.
• New signed Windows installer (mfab command, bundled Python). macOS and Linux install the Python package.
• No automatic updates: install newer versions from the download page.

Traffic between your computers is HTTP on your local network, not encrypted. Use it only on a network you trust.

First-run polish and a package that matches its guide

Includes every change from v0.1.2, with the final documentation and console-state corrections packaged into the signed installer and verified zip.

• The bundled guide now matches the live guide and the behavior in this exact download.
• The model-loading status no longer remains stuck at Starting when the console has already moved to another state.
• The Windows installer is freshly signed, and both current downloads have published SHA-256 fingerprints.

Clean-install fixes across Windows, macOS, and Linux

• doctor reports checks as they finish instead of appearing frozen during a cold dependency import.
• Fixed a false Windows state-directory failure and added the missing huggingface_hub check used by starter-model setup.
• Background startup no longer claims success before the host is actually reachable.
• Added clearer extraction, port, Ctrl+C, and macOS firewall guidance at the point where each issue occurs.
• Suppressed the known LibreSSL compatibility warning without hiding unrelated warnings.

Private chat that works from an empty install

• Added Get a starter model when no model is on disk; the roughly 1 GB download starts only after the user clicks.
• Bring-your-own-key chat can work before a local model is loaded, while local and helper routes remain gated normally.
• Added the on-device privacy meter, answer provenance, and downloadable receipt with no provider key inside.
• Added the mfab CLI to the zip, including doctor, host lifecycle, model, peer, chat, and manual update commands.
• Added clearer capability tiers and first-run guidance, including the guide's What we verify section.

Public-beta foundation

• Signed Windows host/helper role picker, local pairing approval, peer revocation, and manual update checking.
• Local model discovery, loading, chat, and explicit allowed-folder selection through the host console.
• Local usage receipts, capability cards, a guided console tour, and a plain-language user guide.
• Introduced the 2Big2Fit and context retention names as opt-in previews with capacity and continuity boundaries attached.
• Corrected guide encoding and made the update check refuse accidental downgrade prompts.

The build name records release-tree validation work. This public package continued to use the standard local-model path and did not ship the private native acceleration kernels.