Knowledge

Model-infrastructure asset — Quant-variant RAM-floor manifests. the RAM-floor honesty table users choose from. License: license of base (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Model-infrastructure asset — Curated LoRA art packs (consent chain). License: per-LoRA license + artist consent chain — the consent-first packaging rule (see QueHub art doctrine) (license requires per-release verification against upstream before redistribution). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Model-infrastructure asset — Speculative-decoding draft pairs. 2× speed free-lunch. License: inherit base license (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Model-infrastructure asset — Vision projectors for GGUF stacks. multimodal-on-cpu plumbing. License: per-upstream (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Model-infrastructure asset — Tokenizer/chat-template compatibility pack. the GGUF↔chat-template truth table (every bad template is a hallucination factory). License: AP/MIT (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — NVIDIA Aegis classifiers. policy taxonomy for org-configurable moderation. License: AP-class (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — ShieldGemma 4B. small safety judge. License: Gemma terms (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — Llama Guard · 11B Vision. policy-gated content judging, shield-side. License: CL (community) (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — Llama Guard · 8B. policy-gated content judging, shield-side. License: CL (community) (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — Llama Guard · 1B. policy-gated content judging, shield-side. License: CL (community) (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Safety classifier — Llama Protection Guide classifiers. intensity/classifiers for the shields (the weld to sh/*). License: CL (community license (use-at-own-risk terms)). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — KenLM n-grams (museum class). museum class — listed so nobody re-invents the need. License: AP (Apache-2.0). Engine that runs it: kenlm query runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — Hunspell dictionaries (per language). spellchecking everywhere. License: per-dict (many GPL/MPL) (license requires per-release verification against upstream before redistribution). Engine that runs it: hunspell runtime (linked). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — LanguageTool grammar data. grammar at auntie-accuracy. License: LGPL runtime + data licenses (GPL-class (engine/runtime boundary; assets carry their own per-file licenses)). Engine that runs it: LanguageTool server runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — sequence labeling for research pilots · POS. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — sequence labeling for research pilots · NER. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — Stanza packages (70+ langs). 70+ languages UD parsing. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

NLP model or data pack — spaCy md/lg models (per language). grammar/NER the offline way. License: MIT/AP per release (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Translation model — NLLB · 3.3B. 200 langs, labeled NC. License: CC-BY-NC 4.0 (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: CTranslate2 via the python adapter. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Translation model — NLLB · 1.3B. 200 langs, labeled NC. License: CC-BY-NC 4.0 (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: CTranslate2 via the python adapter. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Translation model — NLLB · 600M. 200 langs, labeled NC. License: CC-BY-NC 4.0 (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: CTranslate2 via the python adapter. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Translation model — MADLAD-400 (400 languages). 400-language NMT. License: research terms (NC-check) (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: CTranslate2 via the python adapter. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Translation model — OPUS-MT translation sets (100+ pairs). offline translation for 100+ pairs — the auntie-in-another-country model. License: CC-by-4.0-sources (check per pair) (CC-BY-4.0 attribution per source). Engine that runs it: CTranslate2 via the python adapter. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Audio generation model — MusicGen small (NC data terms). License: (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Video-to-audio model — MMAudio video-to-audio. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Video generation model — video generation: listed-with-license-flags, download-only · HunyuanVideo. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Video generation model — video generation: listed-with-license-flags, download-only · LTX-Video. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Video generation model — video generation: listed-with-license-flags, download-only · AnimateDiff. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image restoration model — Swin2SR restoration. License: (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image restoration model — Real-ESRGAN · x4. the upscaster everyone needs. License: BSD3 (BSD-style). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image restoration model — Real-ESRGAN · x2. the upscaster everyone needs. License: BSD3 (BSD-style). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — SDXL ControlNet set. composition discipline entry. License: per-model (many openrail variants) (license requires per-release verification against upstream before redistribution). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — SD 1.5 legacy ecosystem tier. the LoRA/controlnet ecosystem compatibility tier. License: CL (community license (use-at-own-risk terms)). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — SD3.5 medium (community-license care). License: (license requires per-release verification against upstream before redistribution). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — FLUX.1 dev (NC). License: (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — FLUX.1 schnell. the current open photorealism floor. License: AP (Apache-2.0). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — SDXL Lightning 4-step. instant sketch generation. License: non-commerical-leaning license check (license requires per-release verification against upstream before redistribution). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Image generation model — SDXL base 1.0. the local art engine baseline. License: CreativeML OpenRAIL++-M (CreativeML OpenRAIL++-M). Engine that runs it: stable-diffusion.cpp / diffusers python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Object detector — · D-FINE-R50. fast open detection for warehouse/agriculture pilots. License: AP/MIT (license requires per-release verification against upstream before redistribution). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Object detector — · RT-DETR-R50. fast open detection for warehouse/agriculture pilots. License: AP/MIT (license requires per-release verification against upstream before redistribution). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Florence-2 · Large. dense captioning/detection one backbone. License: MIT (check) (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Florence-2 · Base. dense captioning/detection one backbone. License: MIT (check) (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Segmentation model — SAM2 hiera small. segmentation for the camera-understanding pilots. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Depth model — Marigold depth (NC-labeled). best-in-class depth, labeled NC. License: NC-class (check) (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Depth model — Depth Anything v2 small. monocular depth for scenes/robotics. License: AP (check) (Apache-2.0). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Document-layout model — Docling layout + tables. layout+table→markdown, the document→RAG weld. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — Calamity scanned-book pipeline. scanned-book pipeline. License: GPL runtime + AP models (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — TrOCR · base handwritten. handwriting line recognition. License: MIT-class (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — TrOCR · base printed. handwriting line recognition. License: MIT-class (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — Surya multi-script OCR. modern multi-script OCR. License: GPL + model license check (GPL-class (engine/runtime boundary; assets carry their own per-file licenses)). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — EasyOCR models (80 langs). 80 langs research-grade. License: Apache-ish per-lang (check) (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — PaddleOCR v5. the print-table champion, light. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

OCR model — Tesseract language data (100+). 100-language OCR, CPU-honest. License: per-lang (many Apache/MIT) (Apache-2.0). Engine that runs it: Tesseract runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Pixtral 12B. efficient multimodal, docs+images. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Qwen3-VL · 30B-A3B. the current open multimodal reference. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Qwen3-VL · 8B. the current open multimodal reference. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Qwen2.5-VL GGUF ports · 72B. screenshots→structured data. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Qwen2.5-VL GGUF ports · 7B. screenshots→structured data. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Qwen2.5-VL GGUF ports · 3B. screenshots→structured data. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Vision-language model — Moondream 2B. the 1GB image describer — alt text anywhere. License: AP-adjacent (check v0.5 license: MIT) (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — eSpeak NG voices (last-resort floor). License: GPL (runtime not asset) — the last-resort intelligibility floor (GPL-class (engine/runtime boundary; assets carry their own per-file licenses)). Engine that runs it: eSpeak NG runtime (last-resort intelligibility floor). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speaker diarization model — NeMo diar + Parakeet pipeline. the meetings pipeline. License: AP-class (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speaker diarization model — pyannote speaker diarization. License: who-spoke-what for meetings (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Voice-activity detector — Silero VAD. the voice-activity gate everything speaks through. License: MIT (models PD-ish) (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — F5-TTS. research-grade voice gen, NC-labeled. License: CC-BY-NC (check) (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — MeloTTS zh-en. lightweight bilingual. License: MIT-class (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — Coqui XTTS v2. cloning capability ⇒ consent-gated UI, never silent. License: CC-BY-NC (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — Kokoro 82M. the small near-neural quality jump, EN/JA. License: AP (Apache-2.0). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Text-to-speech voice — Piper voices (per language). the TTS default, 40+ langs, tiny CPU. License: per-voice CC/PD (most MIT/Apache-class) (license requires per-release verification against upstream before redistribution). Engine that runs it: Piper CPU TTS runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — SenseVoice small. zh/en/ja/ko/yue rich transducer. License: AP (Apache-2.0). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — NVIDIA Parakeet TDT 0.6B v2. best EN ASR open. License: CC-BY-4.0-class (check) (CC-BY-4.0 attribution per source). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Moonshine · base. on-device/edge-first EN STT. License: AP (Apache-2.0). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Moonshine · tiny. on-device/edge-first EN STT. License: AP (Apache-2.0). Engine that runs it: ONNX python runtime. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Distil · large-v3. 6× faster EN. License: BSD-style (BSD-style). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Whisper large-v3 turbo. near-large speed at small cost. License: AP (Apache-2.0). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Whisper line · medium. 99-language STT, GGUF-quantized. License: AP (large: MP-license check) (Apache-2.0). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Whisper line · small. 99-language STT, GGUF-quantized. License: AP (large: MP-license check) (Apache-2.0). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Whisper line · base. 99-language STT, GGUF-quantized. License: AP (large: MP-license check) (Apache-2.0). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Speech-to-text model — Whisper line · tiny. 99-language STT, GGUF-quantized. License: AP (large: MP-license check) (Apache-2.0). Engine that runs it: whisper.cpp (GGUF) sidecar. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — Jina Reranker v2. License: (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — MS-Marco MiniLM cross-encoder. cheap, good, everywhere. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — BGE Reranker v2 M3. the workhorse cross-encoder. License: MIT (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — Qwen3 Reranker · 8B. second-pass precision for RAG. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — Qwen3 Reranker · 4B. second-pass precision for RAG. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Reranker — Qwen3 Reranker · 0.6B. second-pass precision for RAG. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Qwen3 Embedding · 8B. the current open MTEB leader line. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Qwen3 Embedding · 4B. the current open MTEB leader line. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Qwen3 Embedding · 0.6B. the current open MTEB leader line. License: AP (Apache-2.0). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — SFR Embedding Mistral. past MTEB #1 lineage. License: license check (license requires per-release verification against upstream before redistribution). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Arctic embed m. MTEB-topping small. License: AP (Apache-2.0). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — E5 · Large. 100-language memory glue. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — E5 · Small. 100-language memory glue. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — E5-Mistral-7B-Instruct. LLM-grade embedding, 24GB+. License: MIT (MIT). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — GTE (Alibaba) · Qwen2-7B. retrieval champions. License: MIT (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — GTE (Alibaba) · Large. retrieval champions. License: MIT (MIT). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Jina embeddings v3. multilingual MTEB elite. License: CC-BY-NC (⇒ never in a bundle, NC) (non-commercial (CC-BY-NC class) — labeled, user-download only, never shipped). Engine that runs it: python adapter sidecar (downloaded runtime, weights on demand). Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — Nomic embed text v1.5. long-context + Matryoshka. License: BSD-style (BSD-style). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — MIXBAI embed large (1024d). strong open EN 1024d. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — BGE (BAAI) · M3 multilingual. the default EN/multilingual families. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — BGE (BAAI) · Base. the default EN/multilingual families. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — BGE (BAAI) · Small. the default EN/multilingual families. License: MIT (MIT). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Embedding model — MiniLM L6 v2 (384d). the zero-budget embedder, 384d. License: AP (Apache-2.0). Engine that runs it: ONNX Runtime CPU via QynLM. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Devstral Small 24B. the terminal-native agent model, SWE-bench honest. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — CodeBeagle (SBINTU). tuned for agentic repair. License: mix (check per size) (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — MagiCoder 7B. self-instruct strength. License: license check (scrutinize) (license requires per-release verification against upstream before redistribution). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Granite Code 3.x. enterprise-language breadth. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Code Llama · 34B. the compatibility tier, deprecated-but-alive. License: CL (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Code Llama · 13B. the compatibility tier, deprecated-but-alive. License: CL (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Code Llama · 7B. the compatibility tier, deprecated-but-alive. License: CL (community license (use-at-own-risk terms)). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — DeepSeek-Coder-V2-Lite. MoE code at 24GB. License: MIT (MIT). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Qwen3 Coder · 480B-A35B. long-context agentic coding (480K marketing — verify). License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Qwen3 Coder · 30B-A3B. long-context agentic coding (480K marketing — verify). License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Qwen Coder · 32B. the everyday local pair-programmer. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Qwen Coder · 7B. the everyday local pair-programmer. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — Qwen Coder · 1.5B. the everyday local pair-programmer. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — StarCoder2 (BigCode) · 15B. fill/complete. License: AP (Tulu-ish data notes check) (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — StarCoder2 (BigCode) · 7B. fill/complete. License: AP (Tulu-ish data notes check) (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Code model — StarCoder2 (BigCode) · 3B. fill/complete. License: AP (Tulu-ish data notes check) (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Chat & general LLM — Granite 3.x 8B long-context. RAG-context champion for pilots' evidence windows. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.

Chat & general LLM — OpenAI gpt-oss 120B. the workstation flagship, MXFP4 quant reality. License: AP (Apache-2.0). Engine that runs it: QuePo (candle GGUF) and llama.cpp-sidecar serving. Download-on-demand: weights are checksum-verified into ~/.qynqueos/models/ and never bundled in the OS image. model ≠ routing: installing this asset never changes which model answers any user.