Every model,
running on your device

Kynora AI ships with curated GGUF and MLX catalogs for generation, vision, embedding, and reranking — plus support for local weights you already have and an optional local Ollama server. Every download goes straight from the model's host (Hugging Face) to your device; Kynora AI never sits in the middle.

Chat & Ask Docs generation models

The model that writes answers — for everyday Chat or grounded Ask Docs Q&A. Most come in both GGUF (via llama.cpp) and MLX (Apple Silicon) formats; a few sizes come in only one. Tap a size to see it on Hugging Face.

Qwen3

Alibaba's latest general-purpose family, including a mixture-of-experts variant.

Qwen2.5

General-purpose instruction-tuned models across a wide size range.

Qwen2.5 Coder

Code-specialized variants for programming questions and drafting.

Llama 3.1 / 3.2 / 3.3

Meta's Llama family, from lightweight on-device sizes up to 70B.

Mistral

Mistral, Mistral Nemo, and Mistral Small instruction-tuned models.

DeepSeek-R1-Distill-Qwen

Reasoning-distilled models for step-by-step, chain-of-thought style answers.

GPT-OSS

OpenAI's open-weight model, with Harmony-format reasoning output support.

20B

Vision-capable generation models

Attach an image in Chat and these models can reason over it alongside your text — most in both GGUF and MLX, a few in only one format.

Qwen3-VL

Qwen's newest vision-language family, including a mixture-of-experts variant.

Qwen2.5-VL

Vision-language models for image understanding and document-style visuals.

3B 7B

Gemma 3

Google's multimodal Gemma models with native image input.

Llama 3.2 Vision

Meta's vision-instruction-tuned Llama variant.

11B

Pixtral

Mistral's vision-language model for image and text reasoning.

12B

Ornith & SmolVLM

A mixture-of-experts vision model and a lightweight vision model for compact devices.

Embedding & reranking models

Embedding models index your documents for hybrid search; an optional reranker re-scores the top matches before they reach the generation model.

BGE

BAAI's general embedding family, plus a matching cross-encoder reranker.

Nomic Embed Text

Long-context text embeddings, including a mixture-of-experts variant.

Qwen3 Embedding & Reranker

Qwen's embedding and reranking models, sized to match your device.

The Model Manager shows device-aware recommendations — memory needed, estimated tokens/sec on your chip, and an Optimal / Slower / Incompatible rating — before you download anything. Rerankers run via GGUF only; there's no MLX rerank runtime. MLX models need Apple Silicon, so Intel Macs are offered GGUF models only.

Not in the catalog? Bring your own.

The built-in catalogs cover most use cases, but Kynora AI doesn't lock you into them.

Import local weights

Already have a GGUF or MLX model downloaded? Import it directly in Model Manager — Kynora AI reads its config.json and folder layout to add it alongside the curated catalog, with the same device-fit and speed estimates.

Not every model out there has been tested against Kynora AI, so we can't guarantee an arbitrary import will run without crashing — if you hit an issue with a specific model, reach out and we'll take a look.

Connect a local Ollama server

On a Mac, advanced users can point Chat or Ask Docs at a local Ollama server instead of the built-in models (default http://localhost:11434). It's off by default. If you turn it on, conversation text, including any document excerpts Ask Docs retrieves, is sent to that server. Embedding and reranking always stay on-device. See how Kynora AI compares with Ollama and LM Studio.

Every download — catalog or your own — goes directly from the model's host (Hugging Face) straight to your device. Kynora AI doesn't sit in the middle of that connection or see the contents of what you download.

Model licenses at a glance

Licenses are as published by each model's provider and may change — always check the model's own license page before commercial or redistribution use. Meta's Llama license and Google's Gemma license both include usage restrictions beyond standard permissive licenses like Apache 2.0 or MIT; they are not OSI-approved open-source licenses.

Qwen3, Qwen2.5, Qwen2.5 CoderApache 2.0
Llama 3.1 / 3.2 / 3.3 (incl. Vision)Llama Community License (custom)
Mistral, Mistral Nemo, Mistral SmallApache 2.0
DeepSeek-R1-Distill-QwenMIT (some GGUF builds list Apache 2.0)
GPT-OSSApache 2.0
Qwen3-VL, Qwen2.5-VL 7BApache 2.0
Qwen2.5-VL 3BSee the model's Hugging Face page
Gemma 3Gemma Terms of Use (custom)
PixtralApache 2.0
OrnithMIT
SmolVLMApache 2.0
BGE (embedding & reranker)MIT
Nomic Embed TextApache 2.0
Qwen3 Embedding & RerankerApache 2.0