Llama 3.2 (3B, 8B, 11B)

Llama 3.2 processes natural language instructions across 8 officially supported languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai) plus additional languages from broader training, maintaining coherence across 128K token context windows. The model uses a decoder-only transformer architecture with instruction-tuning (via unspecified RLHF/SFT methodology) to follow complex multi-turn conversations and adapt responses to user intent. Distributed via Ollama's GGUF quantization format for local or cloud execution with streaming response support.

Llama 3.2 supports structured function calling enabling agents to invoke external tools and APIs by generating schema-compliant function calls. The model was tested with real agent workflows before release (per documentation), supporting tool use as a documented capability. Integration occurs via the Ollama API layer, which accepts tool schemas and returns structured function calls that agents can parse and execute. Supports both local execution (via Ollama CLI/SDK) and cloud execution with managed inference.

Llama 3.2 is accessible via Ollama's HTTP API (localhost:11434/api/chat) and official SDKs for Python and JavaScript/TypeScript, enabling integration into applications regardless of programming language. The API accepts JSON-formatted chat messages and returns streaming or non-streaming responses. SDKs abstract HTTP details and provide language-native interfaces for model invocation, supporting both local and cloud execution.

Llama 3.2 understands code context and supports tool-calling for development-related tasks, enabling integration into development workflows and IDE plugins. The model is integrated into applications like Claude Code, Codex, OpenCode, OpenClaw, and Hermes Agent (per documentation), suggesting capability for code analysis, generation, and tool invocation in development contexts. Tool-calling support enables the model to invoke build systems, linters, or other development tools.

Llama 3.2 executes locally via Ollama's optimized GGUF quantization format, targeting low time-to-first-token (TTFT) and high throughput on consumer and server hardware. The model is distributed in quantized form (1.3GB for 1B variant, 2.0GB for 3B variant) and loads into GPU VRAM for inference. Ollama abstracts hardware optimization across NVIDIA architectures (with specific mention of Blackwell/Vera Rubin acceleration) and provides streaming response support via HTTP API, enabling real-time token-by-token output.

Llama 3.2 is available via Ollama's cloud infrastructure (Ollama Pro/Max tiers) with managed GPU inference, transparent GPU time-based billing, and geographic routing (US primary, EU/Singapore available). The cloud service abstracts hardware provisioning and scaling, supporting concurrent model limits (1 for Free, 3 for Pro, 10 for Max) and session-based usage tracking. Billing is GPU time-based rather than token-based, with weekly/session limits enforced per tier.

Llama 3.2 performs abstractive and extractive summarization across documents up to 128K tokens, leveraging its extended context window to maintain coherence and capture key information from lengthy inputs. The model uses instruction-tuning to follow summarization directives (e.g., 'summarize in 3 bullet points') and is benchmarked against comparable models on summarization tasks. Summarization occurs via standard chat/instruction interface without specialized summarization endpoints.

Llama 3.2 rewrites and reformulates prompts and instructions, transforming user input into optimized versions for downstream tasks. The model is benchmarked on prompt rewriting tasks and uses instruction-tuning to understand rewriting directives (e.g., 'make this prompt more specific', 'simplify this instruction'). Rewriting occurs via standard chat interface without specialized prompt engineering endpoints.

Llama 3.2 retrieves and synthesizes information from long-context inputs to answer questions across 8 officially supported languages plus broader training languages. The model combines instruction-tuning with 128K token context to perform retrieval-augmented reasoning — given a document or knowledge base, it identifies relevant information and generates answers. Retrieval occurs via semantic understanding rather than explicit indexing, making it suitable for RAG pipelines where documents are provided in-context.

Llama 3.2 rewrites text in different styles, tones, and formats (e.g., formal to casual, technical to plain language, long-form to bullet points). The model uses instruction-tuning to understand rewriting directives and applies transformations while preserving semantic meaning. Rewriting occurs via standard chat interface with natural language instructions specifying the desired style or format.

Llama 3.2 1B variant (1.3GB model size) is optimized for personal information management tasks and edge deployment on resource-constrained devices. The 1B model is competitive with other 1-3B parameter models and supports the same instruction-following, tool-calling, and long-context capabilities as larger variants, but with reduced memory footprint and inference latency. Suitable for on-device deployment on laptops, mobile devices, or embedded systems.

Llama 3.2 11B variant provides increased parameter capacity for more complex reasoning, nuanced instruction-following, and higher-quality outputs compared to 3B variant. The 11B model maintains the same 128K context window and instruction-tuning approach as smaller variants, with improved performance on complex tasks. Model size and VRAM requirements for 11B variant are undocumented, but estimated to be 6-8GB+ based on typical quantization ratios.