WizardLM 2 (7B, 8x22B)

Processes multi-turn chat interactions using a standard role/content message format (user/assistant/system roles) with transformer-based attention mechanisms optimized for instruction-following. Maintains conversation context across turns through full context window utilization (32K tokens for 7B, 64K for 8x22B variants), enabling coherent multi-step dialogues without explicit memory management. Implements instruction-tuning via supervised fine-tuning on complex reasoning tasks, allowing the model to follow nuanced user directives and adapt responses based on conversational context.

Executes chain-of-thought reasoning patterns through transformer attention mechanisms trained on complex reasoning tasks, enabling step-by-step problem solving without explicit prompt engineering. The model decomposes multi-step problems by generating intermediate reasoning tokens that guide subsequent token generation, effectively implementing implicit planning through learned reasoning patterns. Supports both explicit reasoning traces (where the model outputs its reasoning steps) and implicit reasoning (where intermediate computations influence final answers), leveraging the instruction-tuned architecture to recognize when problems require decomposition.

Distributes model weights as open-source artifacts through Ollama's package manager, enabling community inspection, fine-tuning, and redistribution. The model is available under an unspecified open-source license (license terms not documented), with 1.1M downloads on Ollama indicating community adoption. Open-source distribution enables researchers and developers to audit model behavior, implement custom quantizations, and fine-tune for domain-specific tasks without proprietary restrictions.

Supports structured function calling through schema-based tool definitions that the model can invoke to extend its capabilities beyond text generation. The model receives a schema describing available tools (functions, parameters, return types) and learns to recognize when a tool invocation is appropriate, generating structured function calls that applications can execute and feed results back into the conversation. This enables agentic workflows where the model acts as a reasoning engine that orchestrates external tools (APIs, databases, code execution) to solve problems iteratively.

Distributes pre-quantized GGUF-format models through Ollama's package manager, enabling single-command local inference without manual quantization or compilation. Models are downloaded as compressed GGUF artifacts (4.1GB for 7B, 80GB for 8x22B) and loaded into memory for inference via Ollama's C++ runtime, which handles GPU acceleration (CUDA/Metal) and CPU fallback automatically. This approach eliminates cloud API dependencies and latency, enabling private inference with full model control and no data transmission to external servers.

Offers three model size variants (7B, 8x22B MoE, 70B) enabling developers to select optimal performance-cost-VRAM tradeoffs for their deployment constraints. The 7B variant provides lightweight inference suitable for resource-constrained environments (laptops, edge devices), while the 8x22B Mixture-of-Experts variant uses sparse activation to achieve 176B effective parameters with lower VRAM than dense 70B models, and the 70B variant provides maximum reasoning capability for compute-rich environments. Developers can benchmark locally and switch variants by changing the model name in API calls (`ollama run wizardlm2:7b` vs. `ollama run wizardlm2:8x22b`).

Generates text incrementally via streaming API endpoints, returning tokens as they are generated rather than buffering the complete response. Ollama's streaming implementation prioritizes low time-to-first-token (TTFT) through optimized KV-cache management and batch processing, enabling responsive user interfaces that display text as it appears. Streaming is supported across all deployment modes (local REST API, Python SDK, JavaScript SDK, cloud API) via standard HTTP chunked transfer encoding or SDK-level streaming callbacks.

Exposes inference capabilities through a standard REST API (POST /api/chat) and language-specific SDKs (Python, JavaScript) that abstract HTTP details and provide idiomatic interfaces. The REST API accepts JSON-formatted chat messages and returns responses in JSON, supporting both buffered and streaming modes via query parameters. SDKs provide type-safe interfaces (Python: `ollama.chat()`, JavaScript: `ollama.chat()`) that handle serialization, streaming callbacks, and error handling, enabling integration into existing Python/Node.js applications without manual HTTP management.

Provides cloud-hosted inference via Ollama Pro ($20/mo) and Max ($100/mo) subscription tiers, where users pay for GPU time rather than tokens. Sessions reset every 5 hours (intra-session) and 7 days (weekly), with concurrency limits (3 concurrent models for Pro, 10 for Max). Cloud inference uses the same REST API and SDKs as local inference, enabling seamless switching between local and cloud deployments by changing the API endpoint and providing an API key. Cloud deployment handles GPU provisioning, scaling, and maintenance automatically.

Generates text in multiple languages through instruction-tuning on multilingual datasets, enabling the model to recognize language context from input and generate responses in the same language. The model supports language switching within conversations (e.g., user asks in Spanish, model responds in Spanish) without explicit language tags or configuration. Specific supported languages not documented — multilingual capability is claimed but language coverage, quality per language, and language-specific limitations are unknown.

Generates responses that incorporate full conversation history up to the context window limit (32K tokens for 7B, 64K for 8x22B), enabling the model to reference previous messages, maintain character consistency, and avoid repeating information. The model processes the entire conversation history as input tokens, using transformer attention to weight recent messages more heavily while still considering earlier context. When conversation history exceeds the context window, the application must implement truncation strategies (e.g., sliding window, summarization) to fit within limits.