DEV.co
Open-Source LLM · swiss-ai

Apertus-70B-Instruct-2509

Apertus-70B-Instruct-2509 is a 70-billion parameter open-source language model from Swiss AI designed for multilingual text generation. It supports over 1,800 languages, respects data owner consent (including retroactive opt-outs), and uses only openly documented training data and methods. The model is available under Apache 2.0 license with no access restrictions, making it suitable for organizations seeking transparent, compliant LLM infrastructure.

Source: HuggingFace — huggingface.co/swiss-ai/Apertus-70B-Instruct-2509
70.6B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
40.4k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerswiss-ai
Parameters70.6B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads40.4k
Likes188
Last updated2026-06-20
Sourceswiss-ai/Apertus-70B-Instruct-2509

What Apertus-70B-Instruct-2509 is

Decoder-only transformer pretrained on 15 trillion tokens using web, code, and math data with a custom xIELU activation function and AdEMAMix optimizer. Post-training included supervised fine-tuning and QRPO alignment. Supports 65,536-token context by default. Trained on 4,096 GH200 GPUs using Megatron-LM. Available in bfloat16 precision. Model weights, training data reconstruction scripts, and intermediate checkpoints are publicly available.

Quickstart

Run Apertus-70B-Instruct-2509 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="swiss-ai/Apertus-70B-Instruct-2509")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Multilingual Customer Support & Localization

Native support for 1,811 languages enables deployment across global markets without separate models. Instruction-tuned variant suits conversational support workflows.

Compliance-Sensitive Applications (EU AI Act, Data Privacy)

Model respects data owner opt-outs (retroactively), includes EU AI Act transparency documentation, and provides contact channels for PII/copyright removal requests. Suitable for regulated industries.

Self-Hosted & Private LLM Deployments

Fully open weights, transparent training, and support across vLLM, SGLang, Transformers, and MLX enable on-premise or private-cloud deployments without closed-source dependencies.

Running & fine-tuning it

70B parameters in bfloat16 precision ≈ 140 GB VRAM minimum for full model inference. Practical deployment on A100 (80GB) or H100 requires gradient checkpointing, LoRA, or quantization (int8/int4 ≈ 35–70 GB). Estimate for single-GPU serving: H100 80GB with batch size 1–2. Multi-GPU / distributed serving (vLLM, SGLang) recommended for production. Training: card lists 4,096 GH200 GPUs; not reproducible at smaller scale without architectural changes.

LoRA fine-tuning is feasible given transformer architecture. QLoRA (quantized LoRA) recommended for single A100 setups. Card mentions post-training included supervised fine-tuning and QRPO; no explicit public recipes or LoRA configs provided. Intermediate training checkpoints available on GitHub (swiss-ai/Apertus-70B-2509 branches), enabling continued pretraining or domain-specific adaptation, though limited official guidance on procedures.

When to avoid it — and what to weigh

  • Need Production-Grade Output Filtering — Card explicitly states no PII output filter is currently provided. Recommended to check quarterly for filter updates; critical for applications handling sensitive user data without additional guardrails.
  • Require High Factual Accuracy Guarantees — Model limitations explicitly state generated content may not be factually accurate and can contain biases from training data. Use as assistive tool only; always verify important information independently.
  • Inference on Consumer GPUs Only — 70B model in bfloat16 requires significant VRAM (~140GB). Feasible only on high-end A100/H100/GH200 hardware or via quantization. Cost-prohibitive for latency-sensitive low-margin applications without optimization.
  • Need Extended Proprietary Lock-In or Vendor Support — Fully open model means no exclusive vendor features or SLA guarantees. Support is community-driven; production deployments require in-house infrastructure expertise.

License & commercial use

Apache 2.0 license. Permissive OSI-approved license allowing modification, distribution, and commercial use with attribution and liability disclaimer. No copyleft restrictions. License is clear and standard.

Apache 2.0 is a permissive open-source license explicitly permitting commercial use, redistribution, and derivative works. No license-level restrictions on commercial applications. However, model developers request contact ([email protected]) for PII/copyright removal and provide contact channels ([email protected], [email protected]) for data protection compliance. Commercial users must implement own PII output filtering and respect copyright removal requests. No proprietary restrictions, but due diligence on training data and legal review recommended for high-stakes applications.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No security audit or threat model documented. Card acknowledges model can generate inaccurate, biased, or harmful content; no built-in content filtering. Explicitly states no PII output filter currently available (recommended quarterly checks). Training data respected opt-out consent and avoids memorization but no formal memorization tests published in provided excerpts. Data subject removal requests handled via email contact; process SLA and audit trail unknown. Deployment security (access control, monitoring) is user responsibility. Use in regulated environments (finance, healthcare, legal) requires additional compliance controls and internal review.

Alternatives to consider

Llama 3.1 (70B)

Similar scale, strong multilingual support, and comparable pretraining benchmarks (67.3% avg vs. Apertus 67.5%). Llama has larger ecosystem and more production deployments; but closed-door training and less explicit data compliance documentation. Choose if ecosystem maturity outweighs openness.

Qwen2.5 (72B)

Highest benchmark average (69.8%) in comparison table. Slightly larger, strong multilingual, native code support. However, Alibaba-backed; less transparent on data compliance and EU AI Act alignment. Choose if pure performance benchmarks are primary driver.

OLMo2 (32B)

Fully open alternative from AI2 with strong transparency (open weights + data). Smaller (32B) but supports many of Apertus' principles (open training, research-focused). Lower absolute performance (67.7% avg) but lower inference cost. Choose if compliance + cost efficiency matter more than scale.

Software development agency

Ship Apertus-70B-Instruct-2509 with senior software developers

Apertus offers transparency and multilingual capability without vendor lock-in. Devco can help you evaluate infrastructure needs, fine-tune for your domain, and design compliant AI workflows. Contact us to assess feasibility for your use case.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Apertus-70B-Instruct-2509 FAQ

Can I use Apertus-70B commercially?
Yes. Apache 2.0 license explicitly permits commercial use without restriction. However, you must implement your own safeguards: no built-in PII filter exists, so output filtering is your responsibility. Respect copyright and data subject removal requests (contact [email protected] or [email protected]). For regulated use (healthcare, finance, legal), conduct independent legal review.
What GPU do I need to run this model?
Full inference in bfloat16 requires ~140 GB VRAM (e.g., H100 80GB or 2x A100 80GB). For single A100, use quantization (int8 ~70GB, int4 ~35GB) or LoRA-based approaches. Inference frameworks vLLM and SGLang support multi-GPU serving and KV-cache optimization to improve throughput. Training from scratch requires 4,096 GH200s; fine-tuning requires far less (single GPU feasible with QLoRA).
How many languages does Apertus support?
1,811 natively supported languages. Substantially broader than most comparable models. Supports long context (up to 65,536 tokens) by default. Multilingual evaluation results on XNLI and XCOPA benchmarks provided in technical report (arxiv:2509.14233).
Is there a production-ready PII / content filter included?
No. Card explicitly states: 'Currently no output filter is provided.' Developers recommend checking the repository quarterly for a future filter reflecting data protection deletion requests. For production use, implement your own output filtering or external content moderation layer.

Software development & web development with DEV.co

Adopting Apertus-70B-Instruct-2509 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy Open, Compliant LLMs?

Apertus offers transparency and multilingual capability without vendor lock-in. Devco can help you evaluate infrastructure needs, fine-tune for your domain, and design compliant AI workflows. Contact us to assess feasibility for your use case.