Open-Weight LLM · Private & Custom AI
Apertus-8B-Instruct-2509
An 8B multilingual instruction model built entirely on open data and compliant training practices—designed for companies running private AI without legal liability or data leakage risk.
Apertus-8B-Instruct is a fully open-weight model (Apache 2.0) trained on 15T tokens with explicit data compliance, supporting 1,811 languages and 65K token context. For ops teams, it's a pragmatic choice: you get transparent training provenance, no closed-door data sources, and the ability to run it entirely within your infrastructure without vendor lock-in or downstream IP concerns.
Model facts
Private deployment
Run Apertus-8B-Instruct-2509 in your own environment
Deploy it on-premises or in a private cloud using standard inference frameworks (vLLM, SGLang, Transformers, or MLX). The 8B variant (~16–24 GB VRAM in bfloat16, ~8–12 GB in quantized form) fits a single GPU or modest CPU inference. All weights, training data reconstruction scripts, and intermediate checkpoints are public—no phone-home telemetry, no third-party data collection. This is the core draw: your proprietary business context stays in your environment, period.
Operational AI use cases
Internal Knowledge Base & Compliance Q&A
Load your company policies, regulatory docs, or SOPs into a private RAG system. Apertus answers questions about internal procedures, compliance frameworks, or product specs without exfiltrating queries to external APIs. Multilingual support lets HR/Legal/Ops serve a global workforce.
Automated Support Ticket Triage & Routing
Run Apertus on incoming tickets (email, chat, forms) to classify severity, extract intent, and suggest routing to the right team. Long context (65K tokens) handles lengthy customer threads. Keep all ticket data and conversational metadata in-house; no third-party SaaS vendor sees customer interactions.
Document Summarization & Contract Review Prep
Batch-process contracts, RFQs, or meeting notes through Apertus to extract key terms, risks, and action items before human review. Compliant training (respects opt-out consent) reduces legal friction if your docs touch sensitive data. Tool-use support enables structured output (JSON tables of obligations, deadlines, etc.).
Custom AI
As a base for custom AI
Apertus serves as a strong foundation for custom fine-tuning on domain-specific tasks—legal assist, technical support, ops automation, multilingual content generation. Its open training recipe and intermediate checkpoints let you replicate or adapt the training process. No licensing friction when you build products on top; Apache 2.0 is commercial-friendly and doesn't require model-weight disclosure in your deliverable.
In the operating system
Where it fits
**Knowledge layer**: embed it in a RAG/search pipeline for internal docs. **Agent layer**: use tool-use capability to orchestrate multi-step workflows (CRM lookups, ticket creation, approval chains). **Workflow layer**: as the backbone of automations (summarization, classification, response generation). In an AI OS, it's the reasoning engine for privacy-critical, compliance-first automations—replacing cloud-dependent APIs.
Data control & security
Self-hosting eliminates data exfiltration to SaaS vendors. Queries, context, and outputs remain on your infrastructure—no vendor training data collection, no third-party logging. The model itself is not a security tool; audit your deployment (network isolation, access controls, inference logging) independently. Apertus's compliant training (respects data removal requests) reduces memorization risk, but always validate that sensitive PII is not in your input corpus.
Hardware footprint
**Estimate (bfloat16)**: ~16 GB VRAM for inference on a single GPU (A40, L40, RTX 6000). **8-bit quantized**: ~8–10 GB. **CPU-only (slower)**: feasible for batch/overnight jobs on multi-socket servers. Context length 65K tokens scales memory linearly (rough rule: +1 GB per 8K tokens of batch context). Pretraining used 4,096 GH200s; inference is vastly cheaper.
Integration
Load via Hugging Face `transformers` (v4.56.0+) or vLLM. Drop it into Python services (FastAPI, Django), containerize for Kubernetes, or run on edge devices via MLX. Chat template is standard; integrate via REST/gRPC. Tool-use schema is available—wire structured outputs into downstream systems (ticketing, CRM, document DBs). Multilingual tokenizer handles non-English prompts natively.
When it's not the right fit
- —You need real-time sub-100ms latency at scale—8B models require optimization and batching; consider smaller quantized variants or distillation.
- —Your domain is extremely specialized (rare medical jargon, proprietary domain lexicon) and you lack a large labeled dataset to fine-tune—base performance may disappoint without adaptation.
- —You need strong factual grounding and zero hallucination tolerance—like all LLMs, Apertus can confabulate; pair it with retrieval and fact-checking, not standalone.
- —Your compliance regime prohibits any model trained on internet-scale data (even with opt-out respect)—custom-trained models on your data only may be required.
Alternatives to consider
Llama 3.1 8B
Comparable general-purpose 8B; strong English, fewer languages. Meta's license is permissive but not fully open-data transparent like Apertus. Easier to fine-tune due to larger community.
Qwen2.5 7B
Slightly smaller, strong multilingual coverage. Chinese-friendly tuning. Apache 2.0 license but less transparency on data compliance and opt-out handling vs. Apertus.
OLMo2 7B
Fully open pretraining data and code; strong STEM. Smaller context, fewer languages. Excellent for transparency; less mature ecosystem than Llama or Qwen.
Related open models
FAQ
Can I run Apertus entirely on-premises, offline?
Yes. Download the weights from Hugging Face once, load them with vLLM or Transformers, and run inference on your hardware. No internet access, no telemetry—your data never leaves your network.
Am I allowed to use Apertus in a commercial product?
Yes. Apache 2.0 license permits commercial use, modification, and distribution. You don't need to open-source your product, but do include a copy of the Apache 2.0 license and disclose that Apertus is a component.
How do I handle a PII removal request if a user's data ends up in Apertus outputs?
Contact swiss-ai at [email protected] to request inclusion in their output filter. Download the filter regularly (every 6 months) and apply it to your deployment to mask PII in model responses. Note: no output filter is currently provided; you may need to implement your own filtering logic.
What's the difference between Apertus-8B and Apertus-8B-Instruct?
The base 8B is pretrained only; Instruct (this variant) includes supervised fine-tuning and QRPO alignment for conversational and instruction-following tasks. Use Instruct for chat/Q&A; use base for continued pretraining or specialized fine-tuning on proprietary data.
Build a Private AI System That's Truly Yours
Apertus is built for companies that need transparency, compliance, and control. LLM.co helps you architect private LLM deployments, fine-tune on your data, and automate ops workflows—all in your environment. Let's design your AI OS.