SeekDeep
Your prompts leave the room.
Cloud-only
Every message hits OpenAI, Anthropic, or some third-party API. Your conversation is now their training data — or their incident log.
Pay per token
Heavy users hit rate limits, hit billing walls, or get throttled when usage spikes. The bot's silent when you need it most.
One model fits none
The same GPT call handles chat, code, image prompts, vision — even though no single model is best at all four.
Run it on your GPU.
Local
Llama-3.1, Qwen-VL, SDXL all run on your card. Zero outbound API calls. Chat, image, archive, search history — all on the box that owns the silicon.
Routed
5 chat roles — default, quality, reasoning, fallback, lightweight. The router picks the right model for the task. Translation goes to Gemma-3n. Code goes to Phi-4.
Aware
Task-LRU swaps models in and out of VRAM as needed. Thrashing warnings before OOM. /gpu watch shows it live.
API calls. Ever.
Simple by construction.
VRAM math.
| ROLE | MODEL | QUANT | VRAM |
|---|---|---|---|
| default_chat | Llama-3.1-8B-Instruct | 4bit · NF4 | 5.1 GB |
| quality_text | Mistral-Nemo-Instruct-2407 | 4bit | 7.2 GB |
| reasoning_code | microsoft/phi-4 | 4bit | 9.1 GB |
| lightweight_chat | gemma-3n-E4B-it | 4bit | 3.1 GB |
| image | Dreamshaper-XL | fp16 | 6.8 GB |
| vision | Qwen2.5-VL-3B-Instruct | fp16 | 3.4 GB |
▸ task-LRU keeps just one chat model resident · pin via LOCAL_*_KEEP_RESIDENT=on
Designed end to end.
Where it's headed.
RECENTLY SHIPPED
- v10.35 · stabilization wave · 565 smoke tests pass
- v10.34 · archive integrity, image reply intent classifier
- v10.31 · InstructPix2Pix + CLIPSeg inpainting
- v10.25 · img2img + upscale
- v10.12 · live GPU/VRAM monitoring
UP NEXT
- Real-ESRGAN upscaler · scaffolded
- TTS voice channel · Piper / XTTS
- Persistent memory · per-user/channel JSON
- SAM + GroundingDINO segmentation
- NSFW gate · CLIP-based classifier
Run it locally.
▸ GPL-2.0 licensed · Windows-first · ~6-minute install