LLaVA (7B, 13B, 34B)

Answers natural language questions about image content by processing images through a CLIP-based vision encoder that extracts visual features, then fuses those embeddings with text prompts through Vicuna's language model decoder. The model performs end-to-end training of both vision and language components, enabling it to ground language understanding in visual context and answer questions requiring spatial reasoning, object identification, and scene understanding.

Generates natural language descriptions and captions of images by encoding visual content through the CLIP vision encoder and decoding it into coherent text via the Vicuna language model. The model learns to summarize visual scenes, identify objects and their relationships, and produce human-readable descriptions without requiring explicit question prompts, making it suitable for batch image annotation and accessibility applications.

Enables complete offline operation by running the entire vision-language model locally without requiring cloud API calls, internet connectivity, or external service dependencies. Once the model is downloaded and Ollama is running, inference can proceed indefinitely without network access, making it suitable for air-gapped environments, mobile deployments, or privacy-critical applications.

Supports analyzing multiple images within a single conversation by passing different images in successive turns, enabling comparative analysis, sequential image understanding, or multi-image reasoning. The model maintains conversation history across turns, allowing users to reference previous images and ask questions that require understanding relationships between multiple images.

Extracts and recognizes text from images using improved visual reasoning capabilities introduced in v1.6, which increased input resolution to 4x more pixels and enhanced OCR-specific training. The CLIP vision encoder captures fine-grained visual details of text characters, and Vicuna decodes these into recognized text strings, enabling document digitization, form processing, and text-in-image extraction without specialized OCR libraries.

Performs logical inference and reasoning about visual content by combining CLIP's visual feature extraction with Vicuna's language reasoning capabilities. The model can answer questions requiring multi-step reasoning about spatial relationships, object interactions, scene composition, and implicit visual knowledge, enabling it to go beyond simple object detection to understand complex visual scenarios and their implications.

Maintains conversational context across multiple turns of image-based questions and answers, enabling users to ask follow-up questions, request clarifications, and build on previous responses. The model uses Vicuna's language model to track conversation history and ground subsequent responses in both the image and prior dialogue, creating a stateful chat experience rather than isolated image-question pairs.

Provides three model size variants (7B, 13B, 34B parameters) optimized for different hardware constraints, enabling deployment on consumer GPUs, enterprise servers, or edge devices. Each variant is distributed through Ollama's model library in a proprietary format (likely GGUF quantization) and can be run locally without cloud dependencies, with inference managed through Ollama's HTTP API, CLI, or language-specific SDKs (Python, JavaScript).

Exposes vision-language inference through Ollama's HTTP REST API endpoints (/api/generate, /api/chat), enabling integration with any HTTP client, web framework, or orchestration tool. The API supports streaming responses, message history for multi-turn conversations, and base64-encoded image input, providing a language-agnostic interface to the vision-language model without requiring language-specific SDKs.

Provides native Python (ollama package) and JavaScript/Node.js (ollama package) SDKs that wrap Ollama's HTTP API, offering idiomatic interfaces for model interaction. The SDKs handle base64 encoding of images, message history management, and streaming response parsing, reducing boilerplate code and enabling developers to integrate vision-language inference with minimal setup.

Processes images at up to 1344x1344 pixels with support for dynamic aspect ratios (672x672, 336x1344, 1344x336) introduced in v1.6, enabling fine-grained visual analysis without image resizing or cropping. The vision encoder adapts to different aspect ratios, preserving visual information in wide, tall, or square images while maintaining computational efficiency through resolution-aware processing.

Generates responses token-by-token with streaming output, enabling real-time display of model output as it is generated rather than waiting for the complete response. Streaming is supported through both Ollama's HTTP API (/api/generate with stream=true) and language-specific SDKs, allowing developers to build responsive UIs that show partial results immediately.