PORTABLEMIND
Current BuildFirmware 1.7.1 · NOW SHIPPING

Firmware 1.7.1, now shipping on every USB

Model Manager, a new Qwen 3.5 roster, multi-modal chat, GPU-accelerated voice, and a fine-tuning pass that dropped first-token latency by an order of magnitude on tight hardware.

Every new PortableMind USB ships with Firmware 1.7.1 preloaded. Voice, SCOUT Vision, Phone Access, the desktop-app launcher, chat history, and AES-secured import/export, all running locally on your machine.

Buy USBSetup guideFull changelog →
What changed in 1.5

New in this build

Firmware 1.7.1 is the current shipping build: in-app Model Manager, Qwen 3.5 roster, multi-modal chat, GPU-accelerated voice on Windows, and a fine-tuning pass that dropped first-token latency by an order of magnitude on tight hardware.

  • Model Manager: download, delete, and manage any Ollama-registry model from inside the app.
  • Qwen 3.5 roster across Fast, Smart Default, Smart Premium, and Scout.
  • Multi-modal chat: images work in any chat turn, not just SCOUT.
  • GPU-accelerated Kokoro TTS on Windows (3 to 6x faster voice replies).
  • Unified voice + chat model residency on tight-VRAM hardware.
  • First-token latency 10 to 30x faster on tight VRAM with the 9B Premium model.
  • Memory decay: taught facts fade if not reinforced, keeping context relevant.
Desktop experience

Launches as a desktop app

PortableMind opens as a dedicated desktop application on both macOS and Windows 10/11. No browser tabs, no terminal commands, just double-click and you're in. Voice, vision, and chat are all accessible from one clean window.

What's included

Firmware 1.7.1 feature set

Firmware 1.7.1 ships with the in-app Model Manager, the full Qwen 3.5 model roster, multi-modal chat (images in any turn, not just SCOUT), GPU-accelerated voice on Windows, and the fine-tuning pass that made first-token latency 10 to 30x faster on tight VRAM. All local, all private. No cloud. No telemetry. Everything runs on your machine.

Your USB, your rules

Model Manager

Download any Ollama-registry model onto the drive from inside the app. Delete models with one click. See disk usage. Get VRAM-aware recommendations for what will actually run on your hardware.

In-app overlay lists every installed model, size, modified date, and which is active.
Streaming progress bar for downloads, one-click delete with confirm.
Curated catalog with a new Light 1.5B preset for very old or very cheap machines.
A whole new model family

Qwen 3.5 Roster

Every preset moved to Qwen 3.5. Better reasoning, tighter answers, faster first-token latency across Fast 2B, Smart Default 4B, Smart Premium 9B, and Scout 4B vision.

Fast, Smart Default, Smart Premium, and Scout all on Qwen 3.5.
Experimental alternative Scout on Gemma3:4B for A/B vision comparisons.
Auto-installed on first launch; existing chat history stays intact.
Images in any turn

Multi-modal Chat

Drop an image into any chat turn and the model reads it. Vision is no longer locked to the SCOUT tab; the whole chat surface is image-aware if the routed model can see.

Drag/drop or paste an image directly into the chat window.
SCOUT tab still exists as a dedicated image workspace.
Router transparently sends images to the vision-capable model.
3 to 6x faster on Windows

GPU-Accelerated Voice

Kokoro TTS moved to DirectML. First launch auto-converts the ONNX opset. Voice replies drop from 8 to 15 seconds down to 2 to 4 seconds on compatible Windows GPUs.

DirectML-accelerated on Windows with a compatible GPU.
Auto-fallback ladder: q8+DML then fp16+DML then q8+CPU.
Piper remains the voice stack on macOS.
No more VRAM swap penalty

Unified Voice + Chat Residency

On cards below 10 GB VRAM, voice reuses whatever model chat is running. Zero swap penalty when you switch between them. On larger GPUs, voice keeps its own dedicated model for parallel throughput.

Automatic based on your real card, no config needed.
Removes the 6 second+ VRAM swap that used to hit voice on tight hardware.
Live re-detection if you plug in an eGPU or update drivers.
Faster, quieter, harder

Fine-Tuning Pass

First-token latency dropped from 60 to 90 seconds down to 3 to 8 seconds on 6 GB VRAM with the 9B Premium model. Memory decays cleanly. Transient inference errors self-recover.

KV cache sized per model tier, no more spillover thrash.
Taught facts decay weekly if not reinforced, protected fact types decay slower.
2 automatic retries with backoff on transient EOF errors.
Still on the drive

Everything from 1.5

Voice mode, SCOUT Vision, Phone Access, the desktop-app launcher, thread history, AES-hardened import/export, and the offline Field Manual all remain intact. Nothing was removed to make room.

Voice mode, SCOUT Vision, Phone Access, desktop-app launcher.
Thread history, AES-256 import/export, Offline Field Manual.
Cross-platform: Windows 10/11 and macOS (Intel + Apple Silicon).
// Live UI

Six screens from the running app

These are stills from the firmware 1.7.1 desktop app, actual frames, not mockups. Click any tile to view full-size.

PortableMind 1.7.1 desktop chat with multi-modal image attach and model pickerChat

Full desktop app, tabs for chat, voice, SCOUT. Drop an image into any chat turn on 1.7.1 and the model reads it, no mode switch required.

PortableMind 1.7.1 voice mode with GPU-accelerated Kokoro TTSVoice

Push-to-talk voice mode with adjustable transcription speed and reply style. Kokoro TTS runs on GPU via DirectML on Windows, 3 to 6x faster than 1.5.

PortableMind 1.7.1 SCOUT Vision on the Qwen 3.5 Scout 4B backboneSCOUT Vision

SCOUT Vision on Qwen 3.5. Identify plants, parse documents, describe scenes, all from the model on your USB. An experimental Gemma3:4B backbone is available for A/B testing.

PortableMind 1.7.1 controls panel with model picker, Model Manager entry, and field manual gridModel Manager

In-app Model Manager (new in 1.6) lists installed models, disk usage, and lets you pull any Ollama-registry model with a live progress bar. Delete gated by confirm.

PortableMind 1.7.1 phone access, QR code, local hotspot pairingPhone Access

Use your phone as a second screen for PortableMind. PC hosts a hotspot, phone scans the QR, still air-gapped from the public internet.

PortableMind 1.7.1 conversation thread with fast first-token latency on Qwen 3.5Faster first token

First-token latency on the Qwen 3.5 9B Premium model on 6 GB VRAM dropped from 60 to 90 seconds down to 3 to 8 seconds in 1.7.1. Every answer is still 100% local.

// Hardening

Built to stay stable offline

Kokoro TTS mutex extended to every streaming path, no more concurrent-inference crashes.
Voice-mode wakeBtn listener tracked properly, no memory leak on re-init.
30 s stream-read timeout on the follow-up LLM stream, no more infinite hangs.
SCOUT findings sorted by timestamp before slicing, most recent images surface first.
Per-model inference tuning (temperature, top_p, repeat_penalty) for premium vs default vs fast.
Voice context window pinned to 4096 (2048 on weak hardware) so KV cache stays small.
Silent temp-file cleanup errors now log a warning instead of accumulating.
Why it matters

Hardening in 1.5 focuses on fewer edge-case failures, faster first responses, improved voice and vision pipelines, and keeping file custody intact with AES protection on packs. Offline stays the default.

How you get it

New customers

New units ship with 1.5

Every PortableMind USB leaving the shop is preloaded with Firmware 1.7.1, voice, SCOUT Vision, phone access, the desktop-app launcher, and all current presets. Plug in and you're on the latest build.

Existing owners

Optional update packs

Existing owners can apply an update pack when ready. Your previous firmware keeps running offline until you switch. No pressure, no forced updates.

Offline-first, with no telemetry

No cloud. No telemetry. Everything runs on your machine. Updates are optional, data stays on your drive, and the tools stay usable when everything around you is offline.