PortableMind Firmware 1.7.1: Every Rough Edge, Sanded Down
Firmware 1.7.1 is shipping. It is the current build on every drive going out the door after early July 2026. If you are on 1.7 already, you can pull it as an OTA update. This is not a features release. 1.7 was the features release: Model Manager, Qwen 3.5, multi-modal chat, GPU TTS. 1.7.1 is the pass that came after, when the team ran the whole thing hard on real cards and wrote down every place it stuttered. Every rough edge got sanded down. First token got faster. Voice mode stopped tripping over itself. Memory stopped hoarding trivia. And the runtime learned to recover from the small hardware weirdness that used to bubble up as an error.
TL;DRFirmware 1.7.1 keeps every 1.7 feature intact and quietly fixes the slow first-token, the voice-mode stalls, the memory bloat, and the flaky-VRAM edge cases. If you bought a drive after early July 2026, this is what is running on it.
First token is now 3 to 8 seconds, not 60 to 90
The single biggest complaint on 1.7 was cold-start latency. Ask a Premium 9B model a question on a 6GB VRAM card, and you were waiting a minute or more before the first word appeared. Once it got going it was fast, but that opening pause felt broken.
The cause turned out to be the KV-cache sizing. Every model tier was reserving the same generous context window, which meant 9B models on smaller cards were spending most of their setup time carving out room they did not need. 1.7.1 tightens the context budget per tier so each model only reserves what it can actually use.
On the same 6GB card, the same question, the same model: first token now lands in 3 to 8 seconds. That is not a marketing number. That is the number you will see.
Voice mode responds like it should
Voice mode on 1.7 had a habit of stalling on the first reply, especially if you talked to it right after startup. Sometimes it would just crash the runtime under concurrent load.
The Kokoro TTS engine was guarding the audio thread with a mutex on one code path but not the others. Every streaming path in 1.7.1 now holds the same lock. No more races, no more silent crashes.
First reply feels snappy. Follow-up replies stream cleanly. On cards that used to fall back to CPU TTS under memory pressure, the health-endpoint mutex fix keeps them on GPU.
The companion stops hoarding trivia
When you teach the companion a fact, that fact used to live forever. Names, preferences, that one thing you mentioned about your uncle in March. All of it sat in the memory graph at equal weight.
1.7.1 adds a slow weekly decay. Facts you reinforce stay sharp. Facts you never come back to fade a little each week, then drop off. Protected fact types (identity, allergies, key relationships) decay much slower and get held longer.
The result is a companion that feels less like a filing cabinet and more like a person who remembers what matters. You do not have to prune anything. It handles itself.
SCOUT sorts by recency, VRAM changes are handled live
Small quality-of-life fixes that add up.
SCOUT vision findings now sort newest first. What you asked about last is what shows first. Previously you were scrolling past week-old scans to find the thing you just uploaded.
On flaky 4GB-VRAM setups, transient EOF errors used to surface as hard failures. 1.7.1 retries them twice with backoff before showing you anything. Most of the time you never see the retry happen.
The routing layer also re-detects VRAM live. Plug in an eGPU dock mid-session, update a graphics driver, and the runtime picks up the new capacity without a restart.
- SCOUT findings sort newest first
- 2 automatic retries with backoff on transient EOF errors
- Live VRAM re-detection, no restart required
- GPU TTS stays on GPU under memory pressure
- Concurrent voice-mode requests no longer crash the runtime
Every 1.6 and 1.7 feature is still here
Nothing from the last two releases got dropped or rearranged. 1.7.1 is additive-only under the hood.
Model Manager still lets you pin, swap, and manage local models. Qwen 3.5 is still the default reasoning model. Multi-modal chat still handles image input alongside text. GPU TTS still runs on capable Windows cards.
Chat History, voice profiles, SCOUT scans, and every companion memory you have already built up carry over untouched. This is a maintenance pass, not a reset.
How to get on 1.7.1
If you are already on 1.7, plug your drive in and the updater will offer 1.7.1. It is an over-the-air update, no reflash required.
If you are on 1.6 or older, the /firmware page walks through the upgrade path. You may have a two-step update depending on where you are starting from.
New drives shipping after early July 2026 arrive on 1.7.1 out of the box. You do not have to do anything.
Ready to run AI offline?
PortableMind is the plug-and-run offline AI USB with three tiers: CORE ($49, Windows, chat), v1.7.1 ($79, voice, multi-modal chat, SCOUT, in-app Model Manager), and MAX-SPEED for power users. No internet, no subscription. Pick the tier that fits your needs.
Conclusion
1.7.1 is a small release on paper and a big one in practice. Nothing new to learn, nothing to reconfigure. Just faster first token, quieter memory, cleaner voice, and a runtime that recovers from the small hardware weirdness that used to trip it up. If you are already on a PortableMind drive, plug it in and let the updater do its work.
Read the 1.7.1 release notesFrequently asked questions
Long-tail answers for the search queries around this topic.
- Do I need to reflash my USB?
- No. If you are on 1.7, 1.7.1 is an OTA update from the updater on the drive. If you are on 1.6 or older, the /firmware page walks through the upgrade path.
- Will this break my Chat History or companion memory?
- No. All existing chat history, voice profiles, SCOUT scans, and companion memories carry over. The new weekly memory decay only affects unreinforced facts going forward, and protected fact types decay much slower.
- What if I actually preferred the old first-token behavior on a specific model?
- You can pin models manually in the Model Manager. If you want a larger context reservation on a specific model tier, pin it and adjust the context setting in the Model Manager UI.
- What changed for macOS users?
- Memory decay applies on macOS the same way. Voice mode still uses Piper on Mac since Kokoro-DirectML is Windows only, so the Kokoro mutex fix does not apply to Mac, but Piper on Mac was already stable under concurrent load.
- Is 1.7.1 the current name for the $79 full build?
- Yes. The $79 full build is now called v1.7.1. The name changed with this release. Every new drive ships as v1.7.1.
- Does 1.7.1 change the CORE tier?
- CORE ($49, Windows, chat only) also picks up the first-token latency fix and the retry logic. It does not use voice or SCOUT, so the Kokoro and vision-related fixes do not apply.
- Does the PRO Navigator get this update?
- The PRO Navigator ($199 GPS/AI flagship) is a separate product with its own firmware track. 1.7.1 is a USB drive firmware release. Navigator updates are handled separately.
- How do I confirm I am on 1.7.1?
- Open the runtime, go to Settings, and check the firmware version at the bottom of the panel. It will read 1.7.1 once the update has applied.
Carson Dresser created the first offline AI USB. He builds from South Florida, where hurricane season makes losing power a lived reality rather than a thought experiment, and he watched cloud AI fail people at the worst possible moments: locked accounts, surprise price hikes, assistants that vanish with the signal. In 2025 he decided intelligence should be something you own, and shipped it. Working alone, he engineered the firmware, tuned the local models, wrote the desktop app, and built the over-the-air update system that delivers cryptographically signed payloads to drives already in the field. He also designed the product, built this website, and runs the store, the kind of range that usually takes a company. Before PortableMind he built ClipStitcher and a stack of practical automations for creators and small operators. Every order that ships and every support email still goes through Carson personally.