Bonus: Top up now and we'll double your first deposit — get x2 credits instantly.
Get Started
Sign Up
Sign Up

Trinity Mini: Self-Hosting & Deployment Guide

Trinity Mini is a 26B-parameter (3B active) sparse mixture-of-experts language model from Arcee AI, optimized for efficient long-context reasoning with low per-token cost.

Browse available models

What is Trinity Mini?

Trinity Mini is a 26B-parameter sparse MoE language model from Arcee AI with about 3B parameters active per token for efficient inference. It is primarily used for reasoning-intensive text generation, such as analytical chat, planning, and complex problem solving, while maintaining strong performance on long-context workloads up to around 131k tokens. It is also applied in function calling and multi-step agent workflows where structured tool use and low latency are important. Trinity Mini is the medium-sized model in Arcee AI’s Trinity open-weight family, sitting between Trinity Nano and larger Trinity variants.

How to run Trinity Mini yourself

Open weights (other), 26.1B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.

Trinity Mini VRAM requirements

Precision / quantizationWeightsFigureSource
BF16 (full precision)52.2 GBEstimate26.1B params × 2 bytes
FP827.8 GBFile sizearcee-ai/Trinity-Mini-FP8-Block
MLX 4-bit15.6 GBFile sizemlx-community/Trinity-Mini-4bit

File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.

GPU requirements: deploy Trinity Mini on your hardware

HardwareBF16 (full precision)FP8MLX 4-bit
NVIDIA H200141 GB1×1×1×
NVIDIA H10080 GB1×1×1×
NVIDIA A100 80GB80 GB1×1×1×
NVIDIA L40S48 GB2×1×1×

GPUs needed per variant.

HardwareMLX 4-bit
Mac, 512 GBM3 UltraFits
Mac, 192 GBM2 UltraFits
Mac, 128 GBM4 Max / M3 MaxFits
Mac, 64 GBM4 Pro / MaxFits

Assumes ~75% of unified memory is usable by the GPU.

HardwareBF16 (full precision)FP8MLX 4-bit
AWS p5.48xlarge8× H100 · 640 GBFitsFitsFits
AWS p4de.24xlarge8× A100 80GB · 640 GBFitsFitsFits
AWS g6e.12xlarge4× L40S · 192 GBFitsFitsFits
Google Cloud a3-highgpu-8g8× H100 · 640 GBFitsFitsFits
Azure ND H100 v58× H100 · 640 GBFitsFitsFits

Whole-instance GPU memory; public instance specs.

Serve it yourself

pip install -U vllm
vllm serve arcee-ai/Trinity-Mini
pip install 'sglang[all]'
python -m sglang.launch_server --model-path arcee-ai/Trinity-Mini
pip install -U mlx-lm
mlx_lm.generate --model mlx-community/Trinity-Mini-4bit --prompt "Hello"

Commands use the official repositories above; vLLM and SGLang expose an OpenAI-compatible endpoint.

Not on LLM.API yet

We do not route this model through our API at the moment. Browse the models you can call today — most workloads have a close match already live.

Browse available models

Providers

Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).

Provider Input Output Cache read /M Latency Throughput Uptime
Arcee AI ~$0.25 ~$0.25 — ~160ms ~120 tps 99.9%
AWS Bedrock (Trinity Mini-equivalent) ~$0.30 ~$0.30 — ~190ms ~150 tps 99.9%
Azure OpenAI (Trinity Mini-equivalent) ~$0.35 ~$0.35 — ~220ms ~100 tps 99.95%
Vertex AI (Trinity Mini-equivalent) ~$0.32 ~$0.32 — ~210ms ~130 tps 99.9%

No hosted provider publishes a price for this exact checkpoint on OpenRouter; self-hosting cost depends on your GPUs. Checked weekly. Last updated 4 Oct 2026.

Try this model

Test Trinity Mini right here — free to start.

Trinity Mini
Hi! Want to test the model?

Suggestions for your first prompt

Not available on LLM API yet

We do not route this model through our API at the moment, so there is no endpoint or code snippet for it yet. Browse the models you can call today — most workloads have a close match already live.

Browse available models

Trinity Mini: Self-Hosting & Deployment Guide in production

What the model is good at.

  • Conversational Chat

    Handles general dialogue and instruction-following tasks as a text-only large language model for interactive chat-based applications.

  • Long-Context Reasoning

    Performs efficient reasoning and generation over long contexts around 128k–131k tokens using a sparse mixture-of-experts architecture.

  • Function Calling

    Supports structured tool and function calling, enabling multi-step agent workflows and schema-based integrations with external systems.

  • Structured Output

    Generates well-structured, machine-readable text such as JSON or classified labels suitable for automation, evaluation, and downstream processing.

  • Multilingual Text

    Processes and generates text in multiple languages, enabling cross-lingual chat, drafting, and localization workflows from a single model.

What you get when you call Trinity Mini: Self-Hosting & Deployment Guide through LLM.API.

  • Unified AI Routing

    Automatically send each request to the optimal model across providers based on latency, quality, and cost. One endpoint, dynamic policies, no code rewrites.

    One endpoint, any model

  • Cost-Aware Orchestration

    Control spend with fine-grained rate limits, model tiering, and smart downgrades. Keep performance high while staying within strict budget and quota constraints.

    Predictable, optimized spend

  • Resilient Fallback Flows

    Design multi-step failover chains across providers so requests keep succeeding through outages, rate limits, or timeouts—without changing your application code.

    Never fail on one model

  • End-to-End Observability

    Inspect tokens, latencies, errors, and provider usage in one place. Quickly debug incidents, tune routing rules, and prove reliability to stakeholders.

    One pane of glass

  • Task-Level Abstractions

    Call high-level tasks like chat, tools, or rerank without vendor-specific boilerplate. Swap models freely while keeping a single, stable application contract.

    Code to tasks, not vendors

  • High-Throughput Batch Jobs

    Process millions of inferences via optimized batching with concurrency control, retries, and partial failure handling built in—no custom job infrastructure required.

    Scale inference, not ops

When to Use — When NOT to Use

Use it if...

  • You need a compact language model suitable for on-device or edge deployments.
  • You need cost-efficient inference for high-volume simple chatbots or assistants.
  • Your use case involves lightweight text classification, tagging, or intent detection pipelines.
  • Your use case involves fine-tuning a small model with domain-specific datasets.
  • You need fast inference for autocomplete, query rewriting, or basic summarization tasks.

Avoid if...

  • You need state-of-the-art reasoning performance on complex, multi-step analytical tasks.
  • Your workload requires handling very long context windows with high factual reliability.
  • You need advanced multimodal capabilities like image understanding or video reasoning.
  • You need best-in-class coding assistance across many languages and large codebases.
  • Your workload requires strong safety guardrails and enterprise-grade compliance guarantees out-of-the-box.

COMMUNITY

What developers say about Trinity Mini

LMArena—Speed—Sources: LMArena, Artificial Analysis

Summarised from public developer discussions and write-ups. Opinions are the sources’, not LLM.API’s.

  • Developers on Hacker News report that Trinity Mini has better latency and throughput than GPT-OSS-20B and appreciate its efficiency relative to its 26B parameter (3B active) MoE architecture.
  • Developers on LinkedIn and Hugging Face describe mixed results, noting the model struggles with complex coding tasks and translation, while showing high error rates for tool calling and structured JSON output.
  • Developers on YouTube indicate the model is capable when running locally on Apple Silicon for RAG tasks, though quality may not yet match established open-source standards in certain domains.

SOURCES

Frequently Asked Questions

  • What is Trinity Mini?

    Trinity Mini is a 26B-parameter sparse mixture-of-experts language model by Arcee AI with about 3B active parameters optimized for efficient reasoning over long contexts.

  • What is the context window of Trinity Mini?

    Trinity Mini supports a context window of approximately 131K tokens, enabling long documents, multi-step workflows, and extended multi-turn conversations.

  • What does Trinity Mini cost to use on LLM.API?

    On LLM.API, Trinity Mini typically follows Arcee AI’s pricing of about $0.04–$0.045 per million input tokens and $0.15 per million output tokens, plus gateway overhead.

  • What is Trinity Mini best suited for?

    Trinity Mini is best for long-context reasoning, structured outputs, tool or function calling, and cost-efficient general-purpose chat and automation agents.

  • Which modalities does Trinity Mini support?

    Trinity Mini is a text-only model that accepts text prompts and returns text completions; it does not natively process images, audio, or video.

  • How fast is Trinity Mini in terms of latency and throughput?

    Thanks to its sparse MoE design, Trinity Mini usually delivers fast token throughput comparable to small dense models while handling significantly longer contexts.

  • How do I call Trinity Mini through LLM.API?

    Set the model identifier to the Trinity Mini slug provided by LLM.API in your completion or chat endpoint call, passing prompts and parameters as usual.

  • How does Trinity Mini compare to larger Trinity models?

    Compared with Trinity Large variants, Trinity Mini is cheaper and lighter with slightly lower peak reasoning quality but similar long-context capabilities for many workloads.

  • What are the main limitations of Trinity Mini?

    Trinity Mini can still hallucinate, lacks up-to-the-minute world knowledge, is not fine-tuned for code to the level of specialist coder models, and is text-only.

  • Does Trinity Mini support function calling and tool use via LLM.API?

    Yes, when used through LLM.API, Trinity Mini can be driven with JSON schemas or tool definitions to perform function calling and multi-step tool-using workflows.

COMPARE

Competitive Models

Get one key to every model

Swap your API key. Keep your code.

Browse available models Talk to sales