Open LLMs/moonshotai

Open-Weight LLM · Private & Custom AI

Moonlight-16B-A3B

A 16B MoE model (3B active) trained with Muon optimizer—optimized for private deployment and efficient custom ops automation without sacrificing reasoning quality.

Moonlight-16B-A3B is a Mixture-of-Expert language model with 16B total parameters but only 3B active per token, trained on 5.7T tokens using the Muon optimizer instead of AdamW. It benchmarks above comparable dense models (Llama 3.2-3B, Qwen2.5-3B) while consuming ~50% fewer training FLOPs. For ops teams, this density-to-performance ratio makes it cost-effective to run privately on modest hardware while maintaining strong instruction-following and reasoning.

16B
Parameters
mit
License (OSI/permissive)
Unknown
Context
44k
Downloads

Model facts

Developermoonshotai
Parameters16B
Context windowUnknown
Licensemit — OSI/permissive
Tasktext-generation
GatedNo
Downloads44k
Likes113
Updated2026-01-30
Sourcemoonshotai/Moonlight-16B-A3B

Private deployment

Run Moonlight-16B-A3B in your own environment

Moonlight is architected like DeepSeek-V3 (MoE), runs on standard transformers/vLLM/SGLang stacks, and has no gating. A company can download the weights, deploy on internal infrastructure (GPU clusters, edge servers, or hybrid setups), and keep all inference data, prompts, and outputs within their network perimeter—critical for regulated industries or sensitive ops workflows.

Operational AI use cases

01

Internal Knowledge Base & FAQ Automation

Index proprietary SOPs, policies, and runbooks. Use Moonlight to power a private chatbot that answers employee questions about benefits, IT procedures, or compliance without data leaving the company. MoE efficiency means you can run it on one GPU instance and serve dozens of concurrent queries.

02

Ops Ticket Triage & Routing

Feed incoming support or incident tickets through Moonlight to classify severity, assign category tags, and suggest runbooks. The model's strong performance on MMLU and BBH (65.2, 42.4 respectively) suggests reliable structured reasoning—useful for deciding 'is this critical?' or 'who owns this?' without manual review.

03

Contract & Document Review Automation

Moonlight can extract key terms, flag risks, and summarize legal/finance documents in a self-hosted environment. Its 8K context window is sufficient for typical single-document review; math capability (45.3 MATH, 81.1 CMath) handles numerical extraction (payment terms, thresholds) without external APIs.

Custom AI

As a base for custom AI

Moonlight serves as a strong base model for fine-tuning custom domain agents. The Muon training approach and released intermediate checkpoints allow teams to continue training on proprietary corpora (ops manuals, customer interactions, internal logs) with reduced compute compared to dense alternatives. Its MoE structure is directly compatible with vLLM/SGLang, enabling fast iteration and A/B testing of specialized versions—e.g., a 'compliance assistant' or 'incident responder' variant—all staying in-house.

In the operating system

Where it fits

In an AI operating system, Moonlight sits at the **reasoning/orchestration layer**—powerful enough to be the backbone of workflow agents that decide what action to take next. It can power knowledge retrieval augmentation (choosing which internal doc to fetch), workflow state machines (understanding context and next steps), and multi-turn task planning. The active-3B efficiency means it can run alongside lightweight embedding and vector DB layers on the same modest infrastructure.

Data control & security

Self-hosting Moonlight means no API calls to third-party LLM providers: prompts, responses, and intermediate reasoning stay within your network and storage. This is an *architectural* choice that reduces data egress risk and eliminates dependence on external services' data policies. However, the model itself (its weights and behavior) carries no intrinsic security properties; you remain responsible for infrastructure hardening, access control, and audit logging.

Hardware footprint

**Estimate (unverified).** MoE inference (only active experts loaded): ~8–12 GB VRAM (bfloat16). Full model load (all experts): ~32–40 GB VRAM (bfloat16). Quantized (int8/GPTQ): 4–8 GB. Batch inference and aggressive quantization can fit on a single 40GB A100 or dual 24GB RTX 6000s. Serving via vLLM with paged attention further reduces peak memory.

Integration

Moonlight is drop-in compatible with transformers, vLLM, and SGLang—no custom serving code needed. Integrate via REST APIs (using vLLM's openai-compatible endpoint), batch processing (for bulk ops triage), or embedded inference (via llama.cpp or similar). Requires trust_remote_code=True due to custom architecture; review the repo before production use. Context window is 8K—plan doc chunking for longer inputs.

When it's not the right fit

  • Your ops workflow requires real-time latency <100ms and you cannot afford GPU infrastructure; Moonlight's MoE routing adds ~10–20ms overhead vs. dense models on CPU.
  • You need reliable long-context reasoning beyond 8K tokens (e.g., analyzing full customer contracts, multi-month logs); context extension would require retraining.
  • Your team lacks internal ML ops expertise to manage model serving, monitoring, and fallback logic; running it privately demands more infrastructure ownership than an API.
  • You need multi-modal reasoning (images, audio) or real-time web search; Moonlight is text-only, base capabilities.

Alternatives to consider

Qwen2.5-3B

Dense 3B baseline, denser than Moonlight's active 3B but easier to run on edge devices; no architecture complexity, lower ops overhead, but ~10–15 points lower on MMLU and math benchmarks.

DeepSeek-V2-Lite (MoE, 2.4B active / 15.3B total)

Same MoE architecture as Moonlight, similar param count, but trained with AdamW on 5.7T tokens. Moonlight claims ~2x sample efficiency; trade-off: DeepSeek-V2-Lite may have more mature production tooling and larger community, Moonlight is newer.

Llama3.2-3B

Meta's dense baseline, well-established ecosystem, no custom code, but trained on only 9T tokens and scores noticeably lower on reasoning benchmarks (54.75 MMLU vs. 70.0). Lower complexity, higher ops simplicity.

FAQ

Can I run Moonlight fully offline on my own servers?

Yes. Download the weights (~32–40 GB full precision), load via transformers or vLLM, and run inference entirely within your network. No phone-home, no API calls required. You manage the infrastructure (GPU, networking, updates).

Is Moonlight safe for commercial use?

MIT license permits commercial use. However, review the model card and code (trust_remote_code) for compliance with your use case. The model is a pretrained LLM with no built-in content filtering; you are responsible for prompt guards and output validation in production.

How do I customize Moonlight for my company's ops tasks?

Fine-tune on your proprietary ops data (tickets, docs, internal Q&A) using the released training code and intermediate checkpoints. Muon's efficiency means fewer GPUs and tokens needed compared to dense alternatives. Test with vLLM to measure inference throughput before deploying to production.

What's the context window, and is it enough for real ops workflows?

8K tokens (~6K–8K words). Sufficient for single policy docs, support tickets, and incident summaries. For longer documents, implement chunking, retrieval-augmented generation (RAG), or hierarchical summarization upstream.

Ready to build private ops AI?

Moonlight is engineered for self-hosted deployment and fine-tuning. LLM.co helps you architect custom AI systems that keep sensitive ops data in-house. Start a private Moonlight instance on your infrastructure, or explore other open-weight alternatives—all integrated into a unified AI operating system.