A personal agent that lives on your phone.
Liquid AI's LFM2.5 models chat, plan, use tools, transcribe meetings and remember things on the device, with no cloud account needed.
On a Nothing Phone 2a (sped up): chat on LFM2.5-350M → a tool request routed to LFM2.5-1.2B → a 5-speaker meeting with owners → a skill created from one sentence → the agent dashboard.
Most "AI assistants" are a thin client over someone else's GPU. LiquidChat shows how far a phone can go on its own with small, fast foundation models:
- Private by default. Chat, tool calls, meeting audio, transcripts, memory and documents stay on the phone. Nothing leaves unless you configure a destination and allow it.
- Fast enough to feel native. LFM2.5-350M answers casual chat with the first token in about a second. LFM2.5-1.2B handles tool calls and planning. A router picks the model for each step.
- It does things. The phone runs 35+ real Android actions (flashlight, alarms, DnD, Bluetooth, calendar, contacts, settings, UI automation) behind a permission gateway.
- Bigger models only when you choose them. For long meetings or hard questions, the free LiquidChat Hugging Face Space runs on your ZeroGPU quota. Anyone who forks the app can do the same.
- You can see every step. Every model call, agent run, tool, MCP call, retrieval and meeting can be exported as OpenTelemetry spans to your own backend.
- Download the APK from the latest release (or
release/LiquidChat.apk) and install it on Android 7+. - Open Models and download LFM2.5-350M and LFM2.5-1.2B-Instruct. The download is about 1 GB on Wi-Fi.
- Start a chat. Try "Turn on the flashlight", "Remember that I prefer meetings at 11 am", or "Search Hugging Face for LFM2 models".
- Optional: open Meetings → ⏺ and record your next meeting.
- Streaming on-device chat with live tokens, tokens/s and time to first token.
- Per-step model routing. Chat stays on LFM2.5-350M. Tools, planning and structured output go to LFM2.5-1.2B-Instruct. With your approval, a single hard reasoning step can go to a cloud model you configured (Anthropic, OpenAI-compatible, Gemini, or the LiquidChat Space). Routing considers battery, heat and RAM.
- Multi-step agent runtime.
- Plans are validated, run through a permission gateway, verified (DnD is read back after enabling it) and retried.
- Runs can be cancelled, and they resume after the app is killed without repeating a send.
- Each step is shown in the chat: the goal, 🔧 tools with their arguments, 🧠 reasoning, ✓/✗ and the result.
- Skills and routines. Five built-in skills: meeting prep, morning brief, focus mode, meeting follow-up and end of day. You can also describe a new one in a sentence. Skills are declarative and cannot run code.
- MCP servers. The read-only Hugging Face MCP server is preinstalled. Add any other server in Agent settings.
- Voice in and out. Speak a request, and have replies read aloud in any voice installed on the phone (Settings → Reply voice), at the speed you choose.
- Record, upload or share audio. A foreground service records 16 kHz audio and transcribes it while the meeting is still going, with Whisper, Parakeet TDT v3 or Moonshine. Language is auto-detected (good with Hindi/English code-switching) and translated to English by default.
- Speakers and owners. Voice fingerprints separate speakers. Names come from introductions. Action-item owners come from who actually said "I'll…". Renaming a speaker updates the transcript, minutes and owners at once, without processing the meeting again.
- Client-ready minutes. Summary, decisions, open questions, next steps and action items with real due dates. Export as PDF, Word, Markdown or text, or email them to attendees.
- High-accuracy mode on the LiquidChat Space (optional, free with your HF token): Silero VAD + WeSpeaker diarization, Whisper per speaker turn, and Gemma 4 12B for the minutes. A 52-minute Hinglish meeting takes about 5 minutes. In this mode the phone only records (no live on-device transcript), so each meeting is processed once. If the Space is unavailable, processing falls back to the phone. Source is in
space/. - Stop at any time. Processing on the phone or on the Space can be stopped from the meeting's status banner; whatever was already transcribed is kept.
- Ask about a meeting. Answers come from the minutes and the passages that match your question.
- "Remember that…" saves a memory immediately. Replies draw on saved memories, similar past conversations, and passages from imported documents and meetings, all embedded on the device with Qwen3-Embedding-0.6B.
- Indexing runs in the background with capped stores and least-recently-used eviction. Deleting a meeting removes it from the index.
- See everything it remembers in Agent → Memory: Saved facts and preferences, the Conversations chat can recall, and the Documents (meetings and files) chat can search, down to each indexed chunk. Any of them can be forgotten.
- The embedding model loads only when it fits next to the chat models and is released when idle, so memory never pushes the app over its RAM budget.
- Tasks: one list for your to-dos and every meeting's action items, with natural-language quick add (
send Acme proposal by Friday !high @Acme). - Email drafter: follow-ups, new emails and replies, with a tone you choose.
- Agent Insights: completion, routing, latency, cost, and Kept on this phone, which counts tokens processed on-device versus sent to the cloud and the cost avoided.
| Role | Model | Where it runs |
|---|---|---|
| Chat | LFM2.5-350M | Phone (always loaded) |
| Tools, planning, structured output | LFM2.5-1.2B-Instruct | Phone (always loaded) |
| Mobile-actions fine-tune | LFM2.5-1.2B-Instruct-mobile-actions | Phone |
| Reasoning, vision, larger chat | LFM2.5-1.2B-Thinking, LFM2-VL-450M, LFM2.5-VL-1.6B, LFM2-2.6B, LFM2-8B-A1B | Phone, loaded on demand |
| Speech to text | Whisper (small/medium), Parakeet TDT v3, Moonshine | Phone |
| Embeddings | Qwen3-Embedding-0.6B | Phone, loaded per use |
| Diarization, long-form ASR, minutes | Silero VAD, WeSpeaker, Whisper large-v3 / Hinglish fine-tune, Gemma 4 12B | LiquidChat Space (optional) |
All on-device models run on the Cactus engine (INT4/INT8). The prompt window is raised to 4,096 tokens.
ModelLifecycleManager keeps the two LFM2.5 models resident and loads everything else on demand. It evicts least-recently-used models to stay within a RAM budget it picks for each device class.
Fine-tuning. The mobile-actions model was trained with LoRA (r=16, α=16) via Unsloth on Hugging Face Jobs, using 31,550 examples. Three Claude Code skills in .claude/skills/ automate the loop from training to the Cactus format to the APK: /unsloth-jobs-training, /lora-to-cactus-hub and /build-apk-liquidchat. Chats can be exported as JSONL datasets to feed the next round.
LFM3 will be added as soon as Cactus can run it.
LiquidChat exports traces and metrics over OTLP/HTTP using @traceverse/genai-otel-instrument-node (docs). Spans follow the OpenTelemetry GenAI semantic conventions, so any OTLP backend can read them. Export is off by default.
| Span | What it covers |
|---|---|
cactus.complete, cactus.embed, cactus.stt.transcribe, cactus.init, cactus.download |
Every on-device model call: tokens, time to first token, decode speed, memory, tool calls |
invoke_agent liquidchat / invoke_agent skill:<name> |
One agent run or skill, the parent of its steps |
execute_tool <tool> |
A tool call through the permission gateway; MCP calls carry mcp.server.name |
retrieve memory, retrieve chat_context, retrieve |
Memory, chat-context and meeting retrieval |
query <store> |
Vector-store searches with rag.result.score_max/min/mean/margin |
hf_space <api> |
Calls to the LiquidChat Space |
meeting.process, meeting.process_space |
The meeting pipeline, on the phone or on the Space |
| Remote model calls | Cloud and Space chat calls, with the same token and cost attributes |
Prompts, replies, tool arguments and results, and transcripts are attached only when Include prompts and replies is on. Error messages follow the same rule; otherwise only the error type is sent.
| Field | What to enter |
|---|---|
| Export telemetry | On to start exporting |
| Collector URL | The OTLP/HTTP base URL. The app adds /v1/traces and /v1/metrics. Use https://; plain http:// is accepted only for a collector on your local network. |
| Headers | Comma-separated key=value pairs (the OTEL_EXPORTER_OTLP_HEADERS format). Stored in Android Keystore-encrypted storage; the field shows only •••• saved afterwards. |
| Include prompts and replies | Adds message text, tool I/O and transcripts to spans |
| Send now / Save | Flush buffered telemetry immediately / apply the settings |
The app also flushes whenever it goes to the background.
Any collector with the OTLP/HTTP receiver works:
# otel-collector config
receivers:
otlp:
protocols:
http:
endpoint: 0.0.0.0:4318
exporters:
otlp/jaeger:
endpoint: jaeger:4317
tls: { insecure: true }
service:
pipelines:
traces: { receivers: [otlp], exporters: [otlp/jaeger] }In the app:
- Collector URL:
https://otel.example.com(orhttp://192.168.1.20:4318on your Wi-Fi) - Headers: whatever your collector requires, e.g.
Authorization=Bearer <token>, or leave empty
Reaching a private collector from a phone on mobile data. We test with a tunnel in front of a reverse proxy:
phone ──https──▶ ngrok tunnel ──▶ nginx :9099 ──▶ HAProxy :443 ──▶ collector :4318 ──▶ Jaeger
If the same proxy also serves a web UI, route the OTLP paths to the collector before the catch-all location /. Otherwise /v1/traces is answered by the UI (typically a redirect to a sign-in page):
location ~ ^/v1/(traces|metrics|logs)$ {
proxy_pass http://127.0.0.1:4318; # or the TLS front door that forwards to the collector
client_max_body_size 32m;
}Check the route with curl -X POST -H 'Content-Type: application/json' -d '{"resourceSpans":[]}' https://<your-host>/v1/traces. It should return 200. Over USB, adb reverse tcp:4318 tcp:4318 lets http://localhost:4318 on the phone reach a collector on your computer.
Arize accepts OTLP/HTTP directly:
| Field | Value |
|---|---|
| Collector URL | https://otlp.arize.com |
| Headers | space_id=<your Arize space ID>,api_key=<your Arize API key> |
| Include prompts and replies | On, if you want the content in Arize |
Both values are in Arize under Settings → API Keys (the space ID is also in the space URL). They are saved encrypted on the phone and never written to logs or spans.
The export SDK handles what Arize needs, so nothing else has to be configured:
- Project. Arize files every trace under a project and rejects traces that name none (
HTTP 400: trace export requires project routing). The SDK putsopeninference.project.name(hereliquidchat) on every trace. To use a different project, addx-project-name=<project>to the headers. - Span kinds. Arize classifies spans by
openinference.span.kind. The SDK sets it on every span next to the GenAI attributes:AGENT,TOOL,LLMfor completions and transcription,EMBEDDING,RETRIEVER, andCHAINfor pipelines and model loads. Without it, Arize shows the span kind as unknown. - Nesting. Model calls nest under the agent run, tool call or meeting pipeline that made them, so a meeting is one trace rather than one per transcription.
To check your keys from a computer before configuring the phone:
curl -s -w ' %{http_code}\n' https://otlp.arize.com/v1/traces \
-H "space_id: $ARIZE_SPACE_ID" -H "api_key: $ARIZE_API_KEY" \
-H 'x-project-name: liquidchat' -H 'Content-Type: application/json' \
-d '{"resourceSpans":[]}'
# → {} 200The app sends to one destination at a time. To fan out to several backends (Arize and your Jaeger, say), point the app at your collector and add an exporter per backend there.
- Only data you mark as public may leave the phone by default. Sending personal data to a cloud model needs your privacy policy to allow it and a first-time approval per provider. Notification content never leaves the phone.
- Anything sensitive asks first, and sending is never "always allowed". Retrieved text can never trigger a tool.
- API keys and telemetry headers live in Keystore-encrypted storage. The agent database is SQLCipher-encrypted.
- The LiquidChat Space keeps nothing and uses only open, ungated models. Its GPU time is charged to the caller's own HF token.
src/
├── agent/ Planner, executor, verifier, state machine, plan validation
├── inference/ Model router, routing policy, cloud providers (incl. LiquidChat Space), cost ledger
├── services/ Model lifecycle (RAM budget, eviction), memory + vector store, HF Space client
├── tools/ 35+ native tools and the permission gateway
├── mcp/ MCP client, server store, tool adapter
├── skills/ Skill registry, generator, validator, built-in skills
├── meetings/ Recorder session, transcriber, speaker ID, minutes, documents, Space pipeline
├── memory/ Retrieval, auto-indexing, chat context
├── observability/ OpenTelemetry export and spans, insights, usage ledger, evaluation set
├── security/ Approvals
├── storage/ Encrypted agent database
└── screens/, components/
space/ The LiquidChat Hugging Face Space (Gradio, ZeroGPU)
android/ Native modules: system controls, accessibility service, audio/meeting docs
docs/ Agent runtime, Android layer and spec
More detail: docs/agent-runtime.md, docs/android-native.md, docs/spec/.
Prerequisites: Node.js 20+, JDK 17+, Android SDK with NDK 27.1.12297006, and ANDROID_HOME set.
npm install
npm run android # debug build on a connected device
npm run check # TypeScript (strict), ESLint, Prettier, Jest
# Release APKs: `full` (sideload, includes UI automation) or `play` (Play Store policy-safe)
cd android && ./gradlew assembleFullRelease -PreactNativeArchitectures=arm64-v8aThe APK is written to android/app/build/outputs/apk/full/release/app-full-release.apk. Release builds are signed with the debug keystore, which is fine for sideloading; add your own keystore before publishing.
CI runs npm run check, debug builds of both flavors and Android lint.
| Version | Highlights |
|---|---|
| v1.3.1 | Memory screen shows recalled conversations and indexed documents, instant speaker renaming, a Stop button for meeting processing, Space mode no longer transcribes on the phone first, chat input stays above the keyboard on Android 15+, reply-voice picker, fix for telemetry stopping after its settings were saved |
| v1.3 | Memory that learns procedures and the user's own words without duplicates, sensitive-memory handling, agent tools matched more strictly, telemetry for retrieval, agent graphs and inference metrics (Arize and any OTLP backend), safer saved export headers |
| v1.2 | LiquidChat Space for meetings and larger models, audio export, Q&A that works on real meetings, visible agent plans, memory and RAG with auto-indexing and limits, a 4K prompt window, the Kept on this phone insight, OpenTelemetry export |
| v1.1 | Streaming chat, task-based routing, Hugging Face MCP, skills from one sentence, speaker identification, upload/share recordings, auto language detection |
| v1.0 | Agent runtime with permissions, meetings, tasks, email drafter, memory, skills |
| v0.x | LoRA-tuned LFM2.5-1.2B turning requests into native Android actions |
- LFM3 on the device, once the Cactus engine supports it.
- iOS parity: native modules and Shortcuts integration.
- Retraining the mobile-actions LoRA on the expanded tool set and on exported real conversations.
Built on Liquid AI LFM models, the Cactus on-device engine, Hugging Face (Hub, Spaces, ZeroGPU, Jobs) and OpenTelemetry.
For research and testing. The model weights are subject to their own licenses, including Liquid AI's LFM license.
Kshitij Thakkar: github.com/Mandark-droid · huggingface.co/kshitijthakkar