TAAFT
Free mode
100% free
Freemium
Free Trial
Deals

Apertus 70B Instruct 2509

Apertus 70B Instruct is a decoder-only transformer pretrained on 15 trillion tokens with a staged curriculum of web, code, and math data. It introduces a novel xIELU activation function and was trained from scratch using the AdEMAMix optimizer on 4096 GH200 GPUs via Megatron-LM. Post-training involved supervised fine-tuning and QRPO alignment. The model natively supports 1811 languages and a 65,536-token context window. It supports tool use and agentic workflows, and is designed to respect data owner opt-out consent while avoiding memorization of training data. All training data, intermediate checkpoints, and training recipes are publicly released, making it fully open in weights, data, and process. On general language understanding benchmarks, the 70B model achieves performance competitive with Llama 3.1 70B and Qwen 2.5 72B. Licensed under Apache 2.0.
Text Gen 7
Released: September 17, 2025

Overview

Apertus 70B Instruct is a fully open 70B-parameter decoder-only language model trained from scratch on 15T tokens using a staged curriculum of web, code, and math data. It natively supports 1811 languages and a 65,536-token context window. Post-trained with SFT and QRPO alignment, it supports tool use and agentic workflows. Weights, training data, and recipes are all publicly released. Licensed Apache 2.0.

About Swiss AI Initiative

The Swiss AI Initiative is the world's largest open science/open source effort for AI foundation models, started in December 2023. Seeded with over 10M GPU hours on the Alps supercomputer and a 20M CHF grant from the ETH Domain, it is the first initiative of the Swiss National AI Institute—a partnership between the ETH AI Center and the EPFL AI Center—leveraging 800+ researchers (70 AI-focused professors) from 10+ Swiss academic institutions.

Industry: Research
Location: Zürich, CH
View Company Profile
Last updated: June 30, 2026
0 AIs selected
Clear selection
#
Name
Task