Open-Weight LLM · Private & Custom AI
Apertus-70B-Instruct-2509
70B multilingual model designed for private deployment with full compliance transparency—built for ops teams needing data-local LLM infrastructure without proprietary lock-in.
Apertus-70B is a fully open-weight, 70.6B-parameter decoder-only transformer pretrained on 15T tokens with explicit data compliance and consent-respecting training. It supports 1,811 languages natively, achieves competitive benchmarks (67.5% avg on standard evals), and ships with complete training recipes and data reconstruction scripts. For operations teams, this means a production-grade LLM you can run entirely within your infrastructure with zero data egress and full auditability.
Model facts
Private deployment
Run Apertus-70B-Instruct-2509 in your own environment
Apertus runs on self-hosted infrastructure via Transformers, vLLM, SGLang, or MLX—all fully open stacks. A 70B model in bfloat16 requires ~140–160 GB VRAM (4× H100 / 8× A100 / proportional cluster setups); smaller deployments use quantization (4-bit/8-bit) to land on 35–40 GB single-GPU. Swiss AI publishes training checkpoints and data-reconstruction code on GitHub, enabling you to audit the entire pipeline. Private deployment is architectural—your documents, customer data, and prompts never leave your boundary.
Operational AI use cases
Internal knowledge-base Q&A and document automation
Ingest your company's private docs (policies, runbooks, product specs) into a vector store; use Apertus-70B as the RAG backbone to answer employee queries on-prem. Multilingual support (1,811 languages) handles global teams; no data leaves your VPC. Reduces support-ticket volume and knowledge-base churn.
Workflow automation: ticket triage, summarization, and routing
Wire Apertus into your ops ticketing system (Jira, ServiceNow) to auto-classify, summarize, and route support/ops tickets. Tool-use support enables it to trigger downstream actions (escalations, webhooks). Self-hosted avoids exposing customer PII to third-party APIs.
Compliance and data-governance audit assistant
Use Apertus as an on-prem agent to scan logs, contracts, and data-inventory files for compliance risks (GDPR, HIPAA, SOX). EU AI Act transparency docs included; Apertus itself trained with opt-out consent and memorization-avoidance—audit-friendly for regulated industries.
Custom AI
As a base for custom AI
Strong foundation for building custom LLM products within your data boundary. Apertus's open architecture, instruction-tuning via supervised fine-tuning + QRPO, and tool-use capability make it straightforward to fine-tune on proprietary operational workflows (e.g., domain-specific triage, custom reasoning patterns). 70B parameter count balances latency and quality for most ops tasks; 8B variant available for edge/cost-constrained deployments. Complete training recipes publicly available—you can reproduce or adapt them.
In the operating system
Where it fits
Core reasoning layer in an AI ops stack: sits below orchestration (n8n, Zapier) and above knowledge/data retrieval. Acts as the 'brain' for agentic workflows—decision-making, document analysis, ticket summarization—while RAG modules (vector DB, BM25) feed it context. In LLM.co's terms: knowledge-to-agent bridge in the execution layer.
Data control & security
Self-hosting Apertus ensures data residency: prompts, documents, and outputs remain on your hardware by design. No cloud logs, no third-party model API calls, no vendor lock-in. Apertus's training included consent-respecting data practices and memorization-mitigation techniques (documented in tech report and EU AI Act summary), reducing downstream PII/copyright leak risk—but you're still responsible for output validation and content filtering at deployment. No built-in PII output filter; Swiss AI notes an optional filter is in development.
Hardware footprint
Estimate for 70.6B parameters in bfloat16: ~140–160 GB VRAM (single 8x H100 / 2x 8xA100 node, or GPU cluster). Quantized (4-bit): ~35–40 GB single A100/H100. Inference latency on modest hardware: 5–15 tokens/sec depending on batch size and context length. Training from scratch: 4,096 GH200s used by Swiss AI; fine-tuning on domain data much lighter.
Integration
Load via Hugging Face Transformers (v4.56.0+), vLLM, or SGLang. Supports chat-template formatting (shown in docs); tool use and long-context (up to 65K tokens) natively supported. Integrate via REST APIs (vLLM server, Ollama, custom FastAPI wrappers) into existing ops systems (Jira, ServiceNow, internal dashboards). Recommended sampling: temperature=0.8, top_p=0.9. Quantize (GPTQ, AWQ) or use bfloat16 depending on latency/throughput needs.
When it's not the right fit
- —Sub-second latency required for real-time user-facing interactions—70B models suit batch/background tasks better than live chat.
- —Hallucination/factuality is mission-critical without human review—Apertus can produce plausible-sounding false content; always verify in compliance/legal use.
- —Your ops environment has <50 GB available GPU VRAM and you cannot quantize—the 8B variant is an alternative, but 70B does not fit consumer hardware.
- —Your team lacks DevOps/MLOps capacity to manage model hosting, scaling, and monitoring—requires Docker, Kubernetes, or managed infrastructure.
Alternatives to consider
Llama 3.1 70B (Meta)
Comparable benchmark performance (67.3%), broader ecosystem support, but trained on proprietary Meta infrastructure; no published data-reconstruction recipes or consent-respecting opt-out claims.
Qwen 2.5 72B (Alibaba)
Strongest benchmark avg (69.8%), good multilingual support, quantized variants available; closed-source training pipeline and China-based infrastructure may conflict with data-residency requirements.
OLMo2 32B (AI2)
Smaller, fully open training details and data, strong compliance focus; 32B may be undersized for complex ops workflows vs. 70B, but fits tighter hardware budgets.
Related open models
FAQ
Can I run Apertus entirely on-premises without cloud APIs?
Yes. Deploy via vLLM, Transformers, or SGLang on your own GPUs or cluster. All dependencies are open-source. No telemetry or cloud callbacks by design. You control the entire inference stack.
Is Apertus licensed for commercial / internal business use?
Yes. Apache 2.0 license permits commercial use, modification, and distribution, provided you retain attribution and license notices. Suitable for internal ops automation and proprietary products built on top. Not gated.
What languages does Apertus support, and is it suitable for non-English ops?
Apertus natively supports 1,811 languages. For multilingual support teams or global ops, it can handle docs, queries, and outputs in most languages without separate models. Review multilingual benchmarks in the tech report for domain-specific performance.
How does Apertus handle privacy and memorization compared to other open models?
Apertus was trained with explicit consent-respecting data practices and memorization-mitigation techniques (documented in the EU AI Act summary and tech report). There is no guarantee against memorization, but the design is more privacy-conscious than many competitors. You remain responsible for validating outputs and applying your own filters.
Build Your Private AI Operating System
Apertus is a foundation for building custom AI and ops automation without leaving your data boundary. LLM.co helps you fine-tune, integrate, and scale private LLM systems for your workflows. Start with Apertus—own your infrastructure.