Meshlit beta 5 · Android and Studio build 45

Experimental Meshlit build 45 for Android, Intel macOS, Windows x64 and Linux amd64.

Downloads

• Android: Meshlit-v2-build45-arm64-v8a-debug.apk — bundled starter model; installed and targeted tests passed on Samsung SM-A207F / Android 11.
• Intel macOS: DMG — bundled Qwen2.5 1.5B CPU runtime; dashboard rendering, 12 HyperL native recipes, authenticated offline generation and unload checks passed on the development Intel Mac. Not an Apple Silicon native build.
• Windows x64: MSI — Studio client preview, no bundled local model/native inference engine. CI contracts and packaging passed; physical Windows installation remains unqualified.
• Linux amd64: DEB — Studio client preview for Debian-family systems, no bundled local model/native inference engine. CI contracts and packaging passed; physical Linux installation remains unqualified.

Changes

Dedicated Models, Routing, Usage and VPN dashboards; smaller Basic settings menus; encrypted device-local usage history, estimated budgets and manual reconciliation; optional OmniRoute/FreeLLMAPI presets; default-off bounded Tailscale status observation on Studio and actual app-network observations on Android. Android network observation cannot identify the VPN provider.

G1 hardening validates durable journal replay and fails closed after ambiguous writes. Android credential transactions are committed before use. G1 coverage remains partial; automatic paid routing is deferred. Budgets are device-local estimates, not provider billing caps.

Samsung testing found and fixed Android filesystem-inspection and hard-link incompatibilities in HyperL dataset handling. Native recipes, encrypted multi-chunk import/verify/rotation/export, private-file permissions, usage reopen and credential checks pass. No paid provider calls or live Tailscale account qualification was performed.

Installation and trust

These packages are experimental: Android is debug-signed; macOS is ad-hoc signed and is not notarized; Windows/Linux do not have verified publisher signatures. Verify SHA256SUMS.txt. Do not disable OS protection globally. Existing data, secrets and passphrases should be backed up securely before testing. Production publisher enrollment/signing and broader clean-device/lifecycle/accessibility acceptance remain open.

Read INSTALL.txt for Intel Mac and DESKTOP-INSTALL.md for cross-platform scope. Matching Eclipse Temurin 21.0.12.1 source and provenance are included beside the installers; Java legal notices are in their runtimes. HyperL retains its separate license.

Validation

Final local Android assemblies, instrumentation assembly, HyperL JVM tests and both lint checks passed (zero lint errors; 401 warnings and 19 hints per flavor). Samsung: 3 targeted tests passed. Nine crawler tests and provenance/document checks passed.

Build provenance

Windows/Linux installers: source 4452c2bc79c96865955b7edd647e5a6a36ce34ce, successful three-platform CI. Android/Intel Mac application source: 5a2b179 (subsequent source changes only corrected Android-specific dataset export; it is excluded from desktop compilation). The release checkpoint includes updated packaging and evidence. This release does not claim full Studio/mobile parity, hard spending guarantees, new cryptography or a research-derived performance breakthrough.

The equivalent unsigned Mac PKG was built and retained locally; only the DMG is uploaded to avoid duplicating the 1.1 GB payload.

Meshlit V2 beta 4 · build 43

Meshlit V2 beta 4 · build 43

Experimental Android build for testing. The arm64 APK is for most current Android phones; x86_64 is for compatible emulators; universal includes both. These APKs are debug-signed and contain the pinned Qwen2.5 0.5B starter model, so downloads are about 540–590 MB.

This update

• Chat voice controls now stay inside Chat. Start and Stop are manual; a live microphone meter and speech status show whether input is detected. An empty 30-second capture reports “No speech heard.”
• Photo selection now shows an inline card. It lists the chosen online vision provider/model when one is configured and uses one “Send photo” action. If no vision model is configured, it explains that and opens provider setup; the selected photo remains when returning to Chat.
• A tool permission no longer forces every ordinary greeting through the small model's tool planner. Explicit tool requests still use the selected grants.
• Updated Android user guide and in-app help.

Tested and limits

Build 43 was installed on a Samsung SM-A207F. The photo card, setup link and return path, manual voice controls, offline pack selection, and no-speech timeout were exercised. A plain “Hi” received a normal greeting on the earlier fixed build; build 43 shares that routing code. Unit tests and Android builds/lint are listed in the PR. No online vision profile is configured on that phone, so image analysis through a provider has not been verified. A spoken Whisper transcript plus audible Piper response still needs a person to speak and listen on a phone. The bundled text model cannot see image pixels and can still refuse benign creative prompts (for example, “sing a song” on Samsung); a larger compatible model or configured provider may be needed for stronger answers. Model token counts and token speed remain unknown when its SDK does not report them. Multi-device heavy-model layer sharding and full desktop/server parity remain separate acceptance work.

User guide · Pull request #4

Meshlit v2.0.0-beta.3 · Android build 42

Meshlit v2 build 42 — experimental Android beta

Install the ARM64 APK on ARM64 phones, the x86_64 APK on an x86_64 emulator, or the universal APK if the ABI is uncertain. This is a debug-signed beta (com.meshlit.v2.debug), not a production or Play Store build. It updates an existing V2 debug install signed with the same certificate; back up local identity and receipts before uninstalling.

Build 42 bundles Qwen2.5 0.5B Instruct instead of the older 135M starter. On Samsung SM-A207F, the app migrated the old bundled file and selected Qwen, then answered a fresh Chat prompt using the local model. The Chat composer now shows its selected access route. The illustrated guide includes that phone capture and separates implemented controls from planned features.

The device and settings UI adds clearer record IDs and added times, pairing/QR diagnostics, radar labels and pulse, agent and MCP management routes, cloud/Hugging Face controls, and safer status wording. Each phone creates its own SSH client key after setup; My SSH key shows its public QR and keeps the private key hidden behind screen-lock verification. The SSH host accepts only explicitly enrolled keys and selected command scopes. The chat microphone remembers a verified offline Whisper/Piper route where those packs are installed.

Physical checks: Samsung loaded Qwen and answered one prompt. A Mac SSH client and Meshlit Studio's SSH transport both received an authenticated status response from the Samsung through an ADB USB tunnel, after a temporary STATUS-only grant was enrolled. The grant and listener were removed after testing. Android V1/V2 build, tests and lint results are recorded in the repository progress log.

Still unverified or incomplete: Xiaomi build 42 installation and phone-to-phone SSH need the Xiaomi back online. Direct Wi-Fi SSH was unreachable in this test network; the USB-tunnel test does not prove LAN reachability. Full offline speech conversation, live camera understanding, two-phone model-layer execution, Internet relay for layers, and complete Studio/server parity remain open. Token speed stays unavailable when the engine supplies no trustworthy token count. Local reward credits are not money or independent proof of work. See docs/PENDING_FEATURES.md and the guide for details.

Use SHA256SUMS to verify the downloaded files. release-info.json records the source commit and APK metadata.

Meshlit Studio build 40.1 — Intel Mac offline preview

Meshlit Studio — experimental Intel Mac desktop build 40.1

This expands the previous thin desktop client with a bundled offline model, an owned CPU inference engine and seven searchable management groups. Choose either the DMG or the PKG; both contain the same app and private Java runtime. No separate JDK, Ollama installation or model download is needed for the starter.

• Qwen2.5 1.5B Q4_K_M starter (about 1.04 GiB) with full checksum validation; verified GGUF import and opt-in Hugging Face discovery/download.
• CPU Auto mode verifies AVX2/FMA/F16C and operating-system vector state, with an SSE4.2 fallback. Context, threads, batches, f16/q8_0 key cache, on-demand loading and idle unloading are configurable.
• Basic/Advanced management menus differ; global, settings, chat, saved-model and saved-node search are local. Web search is a separate opt-in browser handoff.
• Real CPU/RAM/process samples and charts, bounded human terminal, pinned SSH/status-only inbound SSH, optional Ghostty handoff, local cryptography/editor and separately licensed HyperL CPU recipes.
• Shared Android/desktop sharding admission allocates complete automatic/manual whole-layer proportions within estimated per-worker memory limits. A desktop live-worker RPC controller and physical cluster benchmark are still pending.

Validation: 521 JVM test executions in the current result files, 9 crawler cases, both Android Full V1/V2 debug assemblies and lints (zero errors, existing warnings retained). Four matched native Intel cases measured AVX2 at 12.1–17.4 versus SSE4.2 at 2.5–7.1 end-to-end tokens/sec; variation is substantial and no general speed multiplier is claimed. Recovered installer checks are recorded in the accompanying evidence.

The app is ad-hoc signed for integrity; the PKG is unsigned. Neither installer is Apple-notarized. Only this Intel Mac on macOS 15.8.1 was physically checked; macOS 11 is the declared packaging floor, not an independently qualified system. Windows/Linux/HarmonyOS NEXT/iOS, full autonomous desktop agents, firewall/capture, VM/container/Kubernetes and live remote cluster execution remain separate gates. This is not a production certification.

Use INSTALL.txt for setup and SHA256SUMS for asset integrity. Matching Temurin Java sources and licence/provenance notices are included. No Android APK is replaced by this desktop release.

All eight uploaded asset names, sizes, states and GitHub SHA-256 digests match the local files. Application packaging source is 12bd20a; the release source target c1f2d60 includes the later documentation reconciliation and portable test-fixture correction, with unchanged Intel application runtime/model binaries.

Documentation tracking covers 153 Markdown files, current plans/progress and separate workstream gates. Source/review status and other-platform CI remain distinct from the Intel payload qualification.

Meshlit v2.0.0-beta.1 · Many nodes. One mind.

Meshlit v2 beta by IMON

Many nodes. One mind. On-device AI runtime + federated inference cluster for Android.

Experimental Android local AI and agent tools, with opt-in audit telemetry and an experimental phone-cluster layer pipeline. Source commit: ad2d3a86f5060ace52c1b46c7a5501253dcc01f7.

Install the matching build

• Full Beta arm64-v8a APK: normal 64-bit ARM Android phones. Advanced automation/VM/network controls require separate opt-ins and compatible installed capabilities.
• Full Beta x86_64 APK: matching emulator or x86_64 Android device.
• Play Review APKs/AAB: narrower permissions, separate .playreview package, autonomous Accessibility/VPN capture and Termux/SMS controls unavailable. These are debug-signed review artifacts, not Play-approved production releases. The AAB is for review/testing, not direct installation.

All APKs include the real pinned SmolLM2 135M Instruct Q4_K_M starter (105,454,432 bytes, about 101 MiB). This explains the large download. The default SDK backend runs locally; larger models have separate download/import controls. Android API 24 minimum; model/runtime and optional hardware support vary. No 32-bit APK is provided in this beta.

Verify your download with SHA256SUMS. Build signing is for testing. Updates from differently signed versions may fail; export anything important before intentionally reinstalling. Reinstall/data clear requires new agreement. See Settings for offline Terms, Privacy, the full guide and optional permissions.

Changes in this beta

• Compact Gemini-inspired chat, settings, fonts/colors, keyboard-aware composer, Models and Monitor status cards.
• Actual-state startup loader and versioned first-use Terms/Privacy agreement in both channels.
• Opt-in encrypted audit metadata, filters, JSONL/CSV export and optional OTLP/HTTP trace/metric collectors; Grafana/Collector templates included.
• Bounded classic-PCAP import and PCAPdroid consent handoff. The legacy built-in VPN that dropped traffic is disabled in Full; live companion capture remains untested.
• Model management, typed task/agent controls, credentials/cloud tools and experimental native layer-pipeline/checkpoint adapters from the current source.
• Detailed build/agent guide, tutorial, feature map, current evidence and planned runtime/repair work.

Validation and limits

Both Full flavors and Play Review APK/AAB build from the final frozen source. Tests: 8 observability, 65 cloud, 46 network, 260 inference and 130 app tests per Full flavor; zero failures. All three app lint reports have zero errors/fatal findings; warnings remain (see PROGRESS.md). One real API35 x86_64 Android audit instrumentation test passes on the final APK, including actual device/task metadata and encrypted reopening/export. The final emulator Chat, Models and Monitor screens were inspected; the bundled starter was shown loaded and the local runtime ready. A normal cold launch succeeded in 7.406 seconds after Gradle stopped; two launches during concurrent lint hit startup ANRs. Physical startup performance and a visual loader-animation capture remain unverified. Real loopback OTLP/protobuf traces/metrics were received; a hosted Grafana account was not tested.

Physical-phone oversized-model execution, replicated cluster recovery, real rooted VM/VNC and several host integrations still need device proof. AI self-repair, signed auto-update, distro installers and Security Lab tooling are planned, not bundled turnkey capabilities. Vendor SDK development telemetry opt-out remains unresolved; see the privacy policy before testing. Audit history is bounded and best-effort, not a complete or tamper-proof compliance archive.

Play release gates include developer-managed AI reporting/content prevention, private privacy contact, SDK data disclosures, production identity/signing, 16 KiB device/AAB testing and Play review. This beta is not a compliance certification.

Source is Apache-2.0 with preserved third-party notices; model/runtime dependencies retain their own licenses. Stryker source was reviewed but not copied. Report reproducible bugs without credentials, prompts or private exported logs: https://github.com/sabbirimon/meshlit-v2/issues