Arcee AI: Trinity Mini

Trinity Mini implements a 26B-parameter sparse mixture-of-experts (MoE) architecture where only 8 out of 128 experts activate per token, reducing computational overhead while maintaining model capacity. The routing mechanism dynamically selects which expert sub-networks process each token based on learned gating functions, enabling efficient inference at 3B effective parameters. This sparse activation pattern allows the model to maintain reasoning quality across 131k token contexts without proportional compute scaling.

Trinity Mini supports structured function calling through schema-based prompting and response parsing, where the model's expert routing mechanism can specialize certain experts for tool-use reasoning. The model accepts JSON schema definitions of available functions and generates structured tool calls in response, with the sparse MoE architecture potentially allocating specialized experts for function selection and parameter binding tasks. Integration occurs via standard LLM API patterns (OpenRouter) with response parsing for function names and arguments.

Trinity Mini maintains coherent reasoning and context awareness across 131k-token input windows through optimized attention mechanisms and expert routing designed for long-sequence processing. The sparse MoE architecture reduces the quadratic complexity of full attention by limiting expert computation to active pathways, while position embeddings and attention patterns are tuned to preserve semantic relationships across extended contexts. This enables the model to perform multi-document analysis, long-form code understanding, and extended conversation history without context truncation.

Trinity Mini's sparse MoE architecture implements dynamic load balancing across 128 experts to prevent bottlenecks where all tokens route to the same expert subset. The routing mechanism uses learned gating functions that distribute token load probabilistically, with auxiliary loss terms during training that encourage balanced expert utilization. This prevents expert collapse (where most tokens ignore certain experts) and ensures GPU compute is distributed across available hardware, maintaining consistent throughput even under variable input patterns.

Trinity Mini applies sparse MoE routing to code-specific reasoning tasks, where certain experts may specialize in syntax understanding, semantic analysis, and code generation patterns. The model processes code tokens through the full 128-expert pool with 8-expert activation per token, allowing the routing mechanism to select experts optimized for programming language constructs, API patterns, and algorithmic reasoning. This specialization occurs implicitly through training on diverse code datasets without explicit expert assignment.

Trinity Mini maintains coherent multi-turn conversations by preserving conversation history within the 131k-token context window and routing tokens through the sparse MoE architecture in a way that respects conversational continuity. The model processes previous turns as context, with the routing mechanism selecting experts that understand dialogue patterns, user intent tracking, and response consistency. Conversation state is managed entirely through context (no explicit memory store), allowing stateless API calls while maintaining semantic coherence across turns.