Zero telemetry by default  ·  No account required  ·  Runs fully offline

AI that lives entirely
on your device.

Run state-of-the-art language models on Mac, iPhone, and iPad using Apple Silicon. No cloud. No subscription required. Your prompts never leave your device.

Download for iPhone & iPadDownload for macOS
100%On-Device
0Data Sent
14B+Params Supported
120+Tok / sec
LenvX AI
Phi-4 Mini 3.8B
Explain the difference between Swift actors and classes
L
In Swift, actors and classes are both reference types, but actors provide built-in data-race protection. Access to actor state must be awaited from outside contexts.
TTFT 84msSpeed 94 tok/sTokens 312
Ask anything…

Live chat preview

Watch it think in real time.

Streaming tokens, transparent reasoning, and full markdown rendering — all running on your device.

Qwen2.5 7B Instruct 4-bitLoaded
Explain the key differences between Swift actors and classes.
Ask anything — runs entirely on your device
4 096 / 8 192 tokens used  ·  Context 50% full

Everything included

Built for power users.

Every feature you need to run, tune, and interact with AI models — without sending a single character to the cloud.

Apple Silicon

Fully On-Device Inference

MLX-optimised weights run entirely on Apple Silicon. No API key. No cloud round-trip. Streaming tokens at up to 120+ tok/s on M-series chips.

Privacy-first

Zero Data Leaves Your Device

No account. No telemetry by default. Chat content is encrypted at rest with ChaChaPoly and never transmitted.

Multimodal

Vision Models

Attach images directly in chat. VLMs like Qwen2-VL and Llama-3.2-Vision understand and describe your images on-device.

Chain-of-thought

Transparent Reasoning

Watch models think step by step. Reasoning spans from <think> tags are surfaced in a collapsible panel — or silently stripped, your choice.

50+ models

Curated Model Library

Browse verified Hugging Face models filtered for your device. One-tap download with SHA-256 verification, resume support, and atomic install.

Voice AI

Voice Input & Read-Aloud

Dictate messages with on-device speech recognition. Tap the speaker icon to have any reply read back with natural text-to-speech.

Pro

Hybrid Cloud Mode

Point LenvX at any compatible cloud endpoint for optional cloud fallback. Your local models remain the default; cloud is opt-in and per-session.

Files + Vision

Attachments & Document Intel

Drop PDFs, text files, code, or images into chat. LenvX extracts text, runs OCR where needed, and passes it as context to the model.

Auto-save

Persistent Chat History

All conversations are saved automatically with JSON snapshots. Pin up to 3 chats. Search, rename, delete — data stays entirely local.

Library

Saved Outputs Library

One tap to save any assistant reply or tool result to your personal library. Browse, search, and export your best generations any time.

Debug

Live Performance Metrics

Tokens per second, time to first token, context usage ring, thermal state, memory footprint — all surfaced in the Advanced inspector.

Customise

System Prompt Studio

Create named system prompts, set per-conversation defaults, and tune sampling parameters (temperature, top-p, top-k) per chat.

Built-in toolkit

12 AI tools, zero setup.

Tap any tool, fill the form, get a result. Every tool sends output to chat for further refinement — and all results run through your local model.

Rewrite Text

writing

Rewrite a passage in Formal, Casual, Concise, Expanded, or Persuasive style.

Fix Grammar

writing

Correct grammar, spelling, and punctuation while preserving your voice.

Compose Email

writing

Draft a professional email from bullet-point notes with a subject line.

Social Caption

writing

Generate three caption options for Instagram, X, LinkedIn, or Threads.

Summarize Notes

productivity

Turn long articles or meeting notes into short, medium, or long summaries.

Resume Bullets

productivity

Turn job notes into 4–6 impact-driven bullets starting with strong verbs.

Idea Generator

productivity

Brainstorm 5, 10, or 15 distinct ideas with elaboration for any topic.

Explain Code

developer

Get a step-by-step explanation of a snippet with bug and performance notes.

Debug Error

developer

Diagnose any stack trace in Swift, Python, JS, TS, Go, Rust, and more.

Translate

language

Translate text into any language, preserving tone and register.

Extract Actions

analysis

Pull owner-tagged action items from a meeting transcript.

Pros & Cons

analysis

Weigh any decision with 3–5 pros and cons plus a neutral recommendation.

Tool outputs flow directly into chat for further refinement
Save any result to your personal library with one tap

Curated catalog

Top models. One tap to install.

Every model is verified against SHA-256 checksums, filtered for your device's RAM, and pre-configured for MLX-optimised inference.

SmolLM2 1.7B Instruct

SmolLM2 · 1.7B

MLXFast

Ultra-compact model for quick tasks on any Apple Silicon device.

Quant4-bit
Disk1.1 GB
RAM2 GB
Speed180 tok/s
8 192 token context
Available

Phi-4 Mini Instruct

Phi · 3.8B

MLXFast

Microsoft's compact powerhouse — surprisingly strong at code and reasoning.

Quant4-bit
Disk2.3 GB
RAM3.5 GB
Speed120 tok/s
16 384 token context
Available

Gemma 3 4B Instruct

Gemma · 4B

MLX

Google's latest compact model with long context and strong instruction following.

Quant4-bit
Disk2.5 GB
RAM4 GB
Speed105 tok/s
32 768 token context
Available

Qwen2.5 7B Instruct

Qwen · 7B

MLXCoding

Alibaba's balanced model excelling at code, math, and instruction following.

Quant4-bit
Disk4.5 GB
RAM5.5 GB
Speed94 tok/s
32 768 token context
Available

Qwen2-VL 7B Instruct

Qwen2-VL · 7B

MLXVision

Full vision-language model — understands and describes images on-device.

Quant4-bit
Disk4.8 GB
RAM6 GB
Speed72 tok/s
32 768 token context
Available

Llama 3.2 11B Vision

Llama · 11B

MLXVision

Meta's 11B vision model with 128k context and best-in-class image understanding.

Quant4-bit
Disk7.2 GB
RAM8.5 GB
Speed48 tok/s
131 072 token context
Available

Mistral 7B Instruct v0.3

Mistral · 7B

MLX

Proven and efficient — great all-around assistant for everyday tasks.

Quant4-bit
Disk4.4 GB
RAM5.5 GB
Speed92 tok/s
32 768 token context
Available

DeepSeek-R1 7B

DeepSeek · 7B

MLXReasoning

Chain-of-thought reasoning model — shows its thinking before answering.

Quant4-bit
Disk4.6 GB
RAM5.5 GB
Speed86 tok/s
32 768 token context
Available

Qwen2.5 Coder 7B

Qwen · 7B

MLXCoding

Specialised code model — write, explain, debug, and review code on-device.

Quant4-bit
Disk4.5 GB
RAM5.5 GB
Speed91 tok/s
32 768 token context
Available

PaliGemma2 3B

PaliGemma · 3B

MLXVision

Google's compact multimodal model — vision-capable on iPhone.

Quant4-bit
Disk2.2 GB
RAM3 GB
Speed130 tok/s
8 192 token context
Available

Llama 3.3 70B Instruct

Llama · 70B

MLXMac-class

Meta's flagship 70B — requires 64 GB Mac. Best-in-class quality.

Quant4-bit
Disk42 GB
RAM46 GB
Speed18 tok/s
131 072 token context
Available

Qwen3 14B

Qwen · 14B

MLXReasoning

Qwen's latest generation with hybrid reasoning and strong coding skills.

Quant4-bit
Disk9 GB
RAM11 GB
Speed42 tok/s
32 768 token context
Available

Custom Import via Hugging Face

Paste any Hugging Face repo ID or URL. LenvX resolves the revision, checks device compatibility, and validates all files before installing.

org/repo-name

Security & privacy

Privacy isn't a setting.
It's the architecture.

LenvX is built from the ground up on the principle that your data is yours. No server operates LenvX AI. No account system exists. Every design decision defaults toward the minimum necessary data.

Zero telemetry by default

Telemetry is off until you explicitly turn it on. Even when enabled, only anonymised, aggregated counts are sent — never prompts, outputs, or model names.

Encrypted at rest

Every message is encrypted with ChaChaPoly (CryptoKit) using a per-install key stored only in Keychain. Even if the SwiftData store is copied off-device, messages remain unreadable.

Host allowlist enforcement

All outbound requests are validated against a strict allowlist. Only Hugging Face CDN and API endpoints can be contacted — your chat subsystem is network-isolated by design.

Multi-stage download verification

Downloads pass TLS, SHA-256 manifest matching, magic-byte executable rejection, path traversal checking, and atomic install via rename(2) before any model is registered.

App Sandbox, minimal entitlements

App Sandbox is enabled. Only outbound network (HF downloads), speech recognition, and user-selected file access are granted. No full-disk access. No server entitlements.

Log redaction

A built-in Redactor scrubs Bearer tokens, HF tokens, email addresses, home paths, and IPs from every log line. Chat content is never logged at any verbosity level.

7-stage download verification pipeline

Every model file passes through independent validation stages before installation. A single failure quarantines the file and surfaces a clear error.

01

TLS 1.2+

System trust store

02

SHA-256 match

Manifest verified

03

Magic-byte scan

No executables

04

Path validation

No traversal

05

Quarantine stage

Isolated until verified

06

Atomic install

rename(2) only

07

HMAC sidecar

Tamper detection

WCAG AA

Accessibility contrast

App Sandbox

Minimal entitlements

CryptoKit

Apple-native crypto only

Platforms

Native everywhere Apple runs.

One app, three platforms, zero compromise. Each platform gets a layout, download strategy, and model catalog specifically designed for its capabilities.

Mac

Apple Silicon

macOS 14+

  • NavigationSplitView 3-column layout
  • Multi-window support with detachable inspector
  • Full menu bar commands & keyboard shortcuts
  • Command Palette (⌘K)
  • Background downloads continue when minimised
  • Up to 70B parameter models on 64 GB M-class Macs

iPad

Universal

iPadOS 17+

  • Adaptive 2-column NavigationSplitView
  • Hardware keyboard hints & shortcuts
  • Stage Manager & multitasking aware
  • Background download with URLSession
  • Slide-over settings sheet
  • Up to 14B models on M-series iPads

iPhone

Universal

iOS 17+

  • NavigationStack optimised for one-thumb use
  • Sticky composer with safe-area awareness
  • Smart offline mode & Low Data Mode support
  • Cellular download warnings before large transfers
  • Voice dictation & read-aloud
  • Compact 1.7B–3B models fit in iPhone RAM

Pricing

Private AI. Free to start.

All core features — including unlimited chat, the full model library, and all 12 tools — are completely free. Pro unlocks cloud mode and priority access.

Free

$0forever
  • Unlimited on-device chat
  • Full model library access
  • All 12 built-in tools
  • Chat history & pinning
  • Attachments & file analysis
  • Voice input & read-aloud
  • Saved outputs library
  • Zero telemetry
Download Free

Pro

Most popular
$39.99/ year
  • Everything in Free
  • Hybrid cloud mode (bring your own API)
  • Custom API endpoint + secure key storage
  • Priority support
  • Early access to new features
  • Unlimited context pinning
Start Pro

Lifetime

$79.99one time
  • Everything in Pro
  • All future Pro features included
  • Pay once, own forever
  • Priority lifetime support
Buy Lifetime

Payments processed by Apple via In-App Purchase. Cancel any time from App Store subscription settings. Lifetime purchase is a one-time IAP, not a subscription.