Llama 3 (8B, 70B)

Generates contextually coherent multi-turn conversations using a Transformer architecture fine-tuned for instruction-following. The model processes chat messages in role/content JSON format, maintaining dialogue state across up to 8,192 tokens of context. Fine-tuning optimizes for natural dialogue patterns rather than raw text prediction, enabling the model to follow user instructions and maintain conversational coherence across multiple exchanges.

Exposes Llama 3 inference through HTTP endpoints (`/api/chat` and `/api/generate`) that support both streaming and buffered response modes. The Ollama runtime handles model loading, quantization, and GPU memory management transparently, allowing developers to call the model via standard HTTP POST requests with JSON payloads. Streaming responses use server-sent events (SSE) or chunked transfer encoding for real-time token delivery.

Ollama Cloud enforces session timeouts (5-hour limit per session) and weekly usage resets, preventing indefinite resource consumption and enforcing fair-use policies across users. Sessions expire after 5 hours of inactivity or absolute time, and weekly limits reset every 7 days. This pattern is designed for shared cloud infrastructure where per-user resource quotas prevent any single user from monopolizing resources.

Llama 3 has been downloaded 23.5M+ times via Ollama, indicating broad community adoption and implicit validation of model quality and usability. The high download count suggests the model is production-ready and widely trusted, though this is a social signal rather than formal certification. Ollama's model registry includes community ratings, reviews, and usage statistics that help developers assess model reliability.

Provides both instruction-tuned and pre-trained base model variants of Llama 3 (8B and 70B), allowing developers to choose between dialogue-optimized models (`llama3`, `llama3:70b`) and raw foundation models (`llama3:text`, `llama3:70b-text`). The instruct variants are fine-tuned for chat/dialogue tasks, while base variants preserve the original pre-training for tasks requiring raw text generation, completion, or custom fine-tuning.

Offers two distinct parameter counts (8 billion and 70 billion) to balance inference speed, memory footprint, and capability. The 8B variant fits on consumer GPUs and runs faster with lower latency, while the 70B variant provides higher quality outputs at the cost of increased memory and compute requirements. Both variants use the same Transformer architecture and training approach, enabling direct capability/performance comparisons.

Supports both local execution (via Ollama CLI/API on user hardware) and cloud execution (via Ollama Cloud with paid tiers). Cloud deployment uses usage-based billing tied to GPU time, with tier-based concurrency limits (Free=1, Pro=3, Max=10 concurrent requests). Local deployment requires no subscription but demands hardware management; cloud deployment trades hardware costs for operational simplicity and automatic scaling.

Implements OpenAI-compatible chat API (`/api/chat`) that accepts messages with role (user/assistant/system) and content fields in JSON format. The model processes multi-turn conversations by maintaining message history and generating contextually appropriate responses. This pattern enables drop-in compatibility with existing chat application frameworks and libraries designed for OpenAI's API.

Provides a `/api/generate` endpoint for raw text completion tasks, accepting a prompt string and generating continuations without role-based structure. This mode is optimized for tasks like code generation, creative writing, summarization, and other non-dialogue text generation. The model generates tokens sequentially until reaching a stop condition (max tokens, end-of-sequence token, or user-specified stop sequences).

Ollama distributes Llama 3 models in a proprietary quantized format (likely GGUF-based, though not explicitly documented) that abstracts quantization details from users. The runtime automatically selects appropriate quantization levels based on available GPU VRAM and hardware capabilities, handling model loading, memory management, and inference optimization transparently without requiring users to manually download or configure quantized weights.

Ollama provides language-specific bindings and examples for Python, JavaScript/Node.js, and cURL, enabling developers to call Llama 3 inference from their preferred language without implementing HTTP clients from scratch. Each SDK abstracts the REST API details while maintaining the same underlying HTTP interface, allowing polyglot teams to integrate the same model across different services.

Ollama Cloud enforces concurrency limits based on subscription tier (Free=1, Pro=3, Max=10 concurrent requests), queuing requests that exceed the limit with a fixed queue size. Requests beyond the queue capacity are rejected with an error. This pattern prevents resource exhaustion on shared cloud infrastructure while allowing burst traffic up to the queue limit.