PortableMind Firmware 1.7: New Models, Multi-modal Chat
Firmware 1.7 is live. The whole model roster moved to the Qwen 3.5 family, chat is now vision-aware end to end, and voice replies on a supported GPU drop from around ten seconds to two. This is the biggest capability jump since 1.5. Two changes carry most of the weight. New models across every slot. And multi-modal that stops treating vision as a separate room. If you already own a drive, plug it in. The update pulls itself on first launch.
TL;DRFirmware 1.7 rolls the full Qwen 3.5 model roster, opens vision to every chat turn, and moves Kokoro voice to the GPU for 3-6x faster replies.
Every model slot moved to Qwen 3.5
The four-model roster kept its shape but every slot got replaced. Better reasoning, tighter answers, faster first-token latency across the board.
You still pick a lane the same way. The default router will pick for you based on the prompt. What changed is what is on the other side of that pick.
- Fast, 2B parameters. Short answers, instant replies, low RAM.
- Smart Default, 4B parameters. The everyday driver for most chats.
- Smart Premium, 9B parameters. For longer reasoning and denser prompts.
- Scout, 4B parameters with vision. The default multi-modal backbone.
Drop an image into any chat
In 1.5 and 1.6, if you wanted the model to read a picture, you had to switch to the SCOUT tab. That is over.
In 1.7, you can drag an image into the chat window, or paste one from your clipboard, and the model reads it as part of the same turn. If the model routed for that turn cannot see, the router quietly upgrades you to Scout for that message and hands the thread back afterward.
SCOUT is still a dedicated tab if you want a clean vision-only workspace. The point is you no longer have to leave chat to ask about a picture.
Kokoro voice moved to the GPU
Kokoro, the TTS engine behind PortableMind voice replies, now runs on the GPU through DirectML on Windows. First-time launch does a one-shot ONNX opset conversion, then caches the result.
On a machine with a compatible GPU, voice replies drop from eight to fifteen seconds down to two to four seconds. Roughly three to six times faster, depending on the card.
If you do not have a GPU, nothing breaks. Voice still runs on CPU. It is just slower, the way it always was.
Chat and voice stop fighting for VRAM
One of the annoying tax lines in earlier firmware was VRAM swap. Switching from chat to voice unloaded one model and loaded another, and you sat there watching it happen.
1.7 fixes that on smaller GPUs. On any card with under ten gigabytes of VRAM, voice mode reuses whatever model chat is currently using. Zero swap. Zero wait. You just start talking.
On cards with ten gigabytes or more, voice keeps its own dedicated model in memory, since you have the headroom for it. Best of both worlds, decided per machine.
A second Scout, if you want to try it
Alongside the Qwen 3.5 Scout, 1.7 ships an experimental Scout built on Gemma 3 4B. It is off by default. You can toggle it on from settings and A/B test it against the default.
The two backbones read images with slightly different eyes. Neither is strictly better. If you use Scout heavily and want to pressure test which one suits your images, the toggle is there. If you do not care, ignore it and the default Qwen 3.5 Scout runs.
What you need to do
Plug the drive in. On first launch, 1.7 auto-pulls the Qwen 3.5 presets, does the ONNX conversion for GPU voice, and you are done. Existing chats, saved voices, and settings all carry over.
No reinstall. No wipe. No manual model downloads. If you were on 1.6, this is the same one-launch update you are used to.
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
Firmware 1.7 is the biggest capability jump since 1.5. Every model got smarter, voice got fast, and vision stopped living in its own room. Update on first launch and you get the whole thing.
Read the full 1.7 release notesFrequently asked questions
Long-tail answers for the search queries around this topic.
- How do I attach an image to a chat?
- Drag and drop it into the chat window, or paste it from your clipboard. The model reads it as part of the same turn. No mode switch required.
- What is the difference between chat with an image and SCOUT?
- SCOUT is still there as a dedicated vision tab if you want it. The change is that you no longer have to leave chat to ask about a picture. Either surface works.
- Do I need to reinstall my models?
- No. Firmware 1.7 auto-pulls the Qwen 3.5 presets on first launch. Your existing chats, voices, and settings stay put.
- What if I do not have a GPU?
- Voice still works on CPU, it is just slower. Chat still works everywhere. GPU acceleration is a speed bonus, not a requirement.
- What GPUs does the new voice path support?
- Kokoro on GPU runs through DirectML, so any DirectX 12 GPU on Windows works. That covers modern NVIDIA, AMD, and Intel Arc cards.
- Will voice mode fight chat for VRAM?
- Not on smaller cards. On any GPU with under 10 GB VRAM, voice mode reuses whatever model chat is already running, so there is no swap penalty. On bigger cards, voice gets its own dedicated model.
- What is the experimental Scout option?
- A second vision backbone built on Gemma 3 4B. You can toggle it on and A/B test it against the default Qwen 3.5 Scout to see which one reads your images better.
- Does this change the tier lineup?
- No. CORE stays at $49 (Windows, chat only). v1.7.1 is $79 with full voice and vision. MAX-SPEED v1.7.1 is $129 on the 400 MB/s drive.
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.