LiquidAI: LFM2-24B-A2B

Executes inference using a Mixture-of-Experts (MoE) architecture where only 2B of 24B total parameters are active per forward pass, reducing computational cost and latency through sparse gating mechanisms. The model routes input tokens to specialized expert subnetworks based on learned routing weights, enabling efficient deployment on resource-constrained devices while maintaining quality comparable to dense models. This hybrid architecture balances model capacity with inference efficiency through selective expert activation rather than full parameter computation.

Maintains coherent dialogue across multiple turns by processing conversation history as context, enabling the model to track entities, maintain conversational state, and reason about prior exchanges. The model uses standard transformer attention mechanisms to weight relevant historical context, allowing it to reference earlier statements, correct misunderstandings, and build on previous reasoning chains. This capability supports both stateless API calls (where full history is passed each turn) and stateful conversation management patterns.

Generates and completes code across multiple programming languages by predicting syntactically and semantically valid continuations of code snippets. The model uses transformer attention to understand code structure, variable scope, and API patterns from context, enabling both single-line completions and multi-function generation. Supports both inline completion (filling gaps in existing code) and full-function generation from docstrings or type signatures.

Interprets natural language instructions and decomposes complex tasks into subtasks or step-by-step execution plans. The model uses attention mechanisms to identify task constraints, dependencies, and success criteria from instruction text, then generates structured plans or reasoning traces. Supports both implicit task decomposition (reasoning internally) and explicit plan generation (outputting step-by-step instructions for external execution).

Generates text informed by provided context or knowledge documents, using attention mechanisms to ground responses in supplied information rather than relying solely on training data. The model integrates context passages into the attention computation, allowing it to cite sources, synthesize information from multiple documents, and reduce hallucination by constraining generation to supported facts. This capability is commonly used in retrieval-augmented generation (RAG) pipelines where external knowledge is injected into the prompt.

Provides real-time text generation through streaming API endpoints, where tokens are emitted incrementally as they are generated rather than waiting for full response completion. The model uses token-by-token generation with streaming protocols (e.g., Server-Sent Events, WebSocket) to enable low-latency user feedback and progressive response rendering. Supports both buffered (full response at once) and streaming (incremental token) output modes.

Generates text constrained to specific formats or schemas (e.g., JSON, XML, CSV, function calls) by using prompt engineering, output validation, or constrained decoding techniques. The model learns to follow format specifications from examples or explicit instructions, enabling reliable extraction of structured data from unstructured prompts. Supports both soft constraints (instructions in prompt) and hard constraints (validation/filtering of generated tokens).

Generates and translates text across multiple languages by routing language-specific tokens through specialized expert pathways in the MoE architecture. The model learns language-specific patterns and vocabulary during training, enabling both translation (source-to-target language conversion) and code-switching (mixing languages in single response). Supports both explicit translation prompts and implicit multilingual generation based on input language.

Adapts model behavior to new tasks or domains by providing examples (few-shot prompting) or task-specific instructions without retraining. The model uses attention mechanisms to learn patterns from provided examples, enabling rapid task adaptation for classification, extraction, summarization, or generation tasks. Supports both explicit examples (few-shot) and implicit adaptation through system prompts or role-playing instructions.