Device-aware recommendations
Review a curated catalog of four instruction-tuned GGUF models below 1B parameters with source revision, license, size, and SHA-256 details.
Cinqic flagship software
Your AI.
Your models.
Your machine.
Juniper is a local-first AI app for people who want a thoughtful assistant, visible model controls, and a clear boundary around private data.
Published alpha prerelease: Juniper 0.4.0-Alpha
01 — Model control
Choose a model without becoming a systems administrator.
Juniper’s Models Market detects the device profile, explains which compact models fit, and manages verified downloads and storage.
Verified model lifecycle
Downloads are resumable and integrity-checked, with atomic installation, pause/resume, and removal.
Model-agnostic by design
No model family is required and no model is bundled in the installer. Compatible providers and runtimes remain visible choices.
Juniper-owned desktop runtime
The desktop bundle owns its loopback llama.cpp process, so ordinary local use does not require Ollama or an account.
02 — Assistant workspace
A useful assistant with the controls exposed.
The chat workspace supports streaming, Markdown, export, private chats, and mobile layouts. Assistant profiles keep personality, model choice, tool policy, memory policy, import, and export in one place.
Optional provider connections include Ollama, OpenAI-compatible servers, and llama.cpp-compatible servers for advanced users and existing setups.
Deterministic composition
System prompt, assistant settings, memories, enabled tools, and recent conversation stay within a context budget.
Explicit permissions
Host-authored calculator, conversion, datetime, attachment, memory, search, and system-info contracts.
Local persistence
SQLite application state, scoped text attachments, diagnostics, and private-chat boundaries.
03 — Data boundaries
The app should tell you where your prompt goes.
Juniper labels routes as ON DEVICE, LOCAL NETWORK, REMOTE, or UNKNOWN, and never treats UNKNOWN as safe. Desktop provider credentials use the operating-system credential store and Android provider credentials use Android Keystore; secrets are not written to SQLite or exports.
Telemetry, analytics, crash reporting, and accounts are not part of this release. A tool the request did not enable is denied rather than executed.
04 — Interface
Conversation first.
Juniper 0.4.0-Alpha has four destinations — Chats, Models, Settings, and a new Agent section in Beta. Reasoning, tool activity, and model details stay one tap away instead of on screen, and the model pill shows whether a reply runs ON DEVICE, LOCAL NETWORK, or REMOTE. The screenshots below are from 0.3.0-rc.33; the conversation and Models screens are unchanged in layout.
Appearance is a set of validated settings: system, light, or dark; standard or high contrast; curated, neutral, or custom accent colors with readable text shades computed for you; bundled Inter, Atkinson Hyperlegible Next, and OpenDyslexic fonts; and density, conversation width, message style, text size, and motion. Nothing is downloaded and no custom code is loaded.
05 — Published release
Juniper 0.4.0-Alpha is available for evaluation.
This is the latest public Juniper release verified on GitHub (machine version 0.4.0-alpha.1), and it matches the current source version. It is an alpha prerelease, not a stable 0.4.0 release.
0.4.0-Alpha adds an Agent section in Beta: give it a goal and it works through it with the folders and tools you allow, previews every file change as a diff, asks before acting, and marks a task Needs review when an answer claims a change that no tool made. Model downloads are now resumable, verifiable multi-file bundles with no 2 GiB cap, and license terms are recorded by the native host. The catalog lists 23 models; Llama 3.x and Gemma 3 entries are held from this alpha because their GGUF files come from third-party mirrors. Full gates and evidence are in the readiness review linked from the release.
Download a platform build
Use the release assets below and verify them against the accompanying SHA-256 manifest.
06 — Known limitations
The boundaries are part of the release.
Linux (.deb, .AppImage) is supported for this alpha; the Windows MSI is supported but unsigned, so SmartScreen warns, and the Agent section is not verified on Windows, where sandboxed code is unavailable. Android stays in Beta: native local inference runs through an in-process llama.cpp bridge for SHA-256-verified GGUF models on ARM64, and a physical ARM64 run is still required before it leaves Beta. The Agent is Beta: its unperformed-claim check is a text heuristic, MCP support is tools-only and local MCP servers are not sandboxed, and backup restore has not been verified on a clean machine. Device Link and Juniper Network remain non-networked protocol and policy work. iOS and macOS are not built.