Current build│
Firmware 1.7.1
Every rough edge from 1.7, sanded down. Faster first token, quieter memory, harder failure recovery. The current shipping build.
// Highlights
Faster First Token
Chat with the 9B Premium model on a 6 GB card, first-token latency dropped from 60 to 90 seconds down to 3 to 8 seconds. KV cache sized per model tier.
Quieter Memory
Taught facts now decay a little each week if not reinforced. Protected fact types decay slower. Stale trivia falls off before it crowds out useful context.
Harder Recovery
Transient EOF errors on 4 GB VRAM get 2 automatic retries with progressive backoff. Kokoro TTS mutex extended to every streaming path.
Live VRAM Re-Route
Plug in an eGPU dock or swap a driver mid-session; the routing layer re-detects VRAM and picks a better model without a restart.
// What changed
- First-token latency down 10 to 30x on tight VRAM (KV-cache sized per model tier).
- Voice-mode TTS mutex now covers every streaming path, no more concurrent-inference crashes.
- Memory: exponential decay per week for taught facts, auto-cull below relevance threshold.
- SCOUT findings now sort by recency before slicing, most recent images surface first.
- EOF retry with progressive backoff on flaky low-VRAM inference.
- Live VRAM re-detection via the hardware profiler event, adapts to eGPU/driver changes.
- Explicit think:false on Qwen chat calls, no more silently eaten num_predict budget.
- Agent-mode context window raised to 16k on strong-tier hardware for multi-step chains.
// Hardening
- Kokoro TTS health-endpoint mutex fix (no more crashes during active inference).
- Voice-mode wakeBtn listener tracked properly, no memory leak on re-init.
- 30 s stream-read timeout on follow-up LLM stream (no more infinite hangs).
- Silent temp-file cleanup errors now log a warning instead of accumulating.
- 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.
← Previous
Firmware 1.7
Whole new model family, and images now work in every mode, not just SCOUT. GPU-accelerated voice on Windows.