PORTABLEMIND
Legacy build│

Firmware 1.7

Whole new model family, and images now work in every mode, not just SCOUT. GPU-accelerated voice on Windows.

[ ORDER PORTABLEMIND ]All versions →
// Highlights

Qwen 3.5 Roster

Full model swap to Qwen 3.5. Fast 2B, Smart Default 4B, Smart Premium 9B, Scout 4B (vision). Better reasoning, tighter answers, faster first token.

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.

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 (3 to 6x faster) on compatible Windows GPUs.

Unified Residency

On cards below 10 GB VRAM, voice reuses whatever model chat is running. Zero swap penalty. Big GPUs (10 GB+) keep voice on a dedicated model.

// What changed
  • Full model roster on Qwen 3.5: Fast 2B, Smart Default 4B, Smart Premium 9B, Scout 4B vision.
  • Multi-modal chat: attach images to any chat turn, not just SCOUT.
  • Experimental alternative Scout on Gemma3:4B for A/B vision testing.
  • Kokoro TTS on GPU via DirectML with auto opset 17 to 22 conversion at first launch.
  • Unified voice + chat residency on tight-VRAM hardware, eliminates 6 s+ swap penalty.
  • Dual-resident voice model on 10 GB+ VRAM for concurrent chat + voice throughput.
  • Per-model num_gpu and num_ctx tuning based on real VRAM and model weights.
// Hardening
  • Auto-detection of VRAM via runtime.json, medic endpoint, and hardware-profiler event (takes max).
  • Kokoro GPU vs CPU fallback ladder: q8+DML then fp16+DML then q8+CPU, with per-strategy validation.
  • First-turn health probe fires immediately on voice mode activation (no cold-start penalty).
← Previous
Firmware 1.6
The drive opens up. Download any Ollama-registry model with a progress bar, delete any model with one click, get VRAM-aware recommendations. Your USB, your rules.
Next →
Firmware 1.7.1
Every rough edge from 1.7, sanded down. Faster first token, quieter memory, harder failure recovery. The current shipping build.