Skip to main content
Language model

Apertus 8B

swiss-ai

Released
September 2, 2025
Data date
September 26, 2026

Apertus 8B is the smaller instruction checkpoint in Swiss AI’s fully open family. Like Apertus 70B, it is described as using 15 trillion pretraining tokens, openly published training resources, and 1,811 natively supported languages. The checkpoint has 8 billion parameters and a 65,536-token context.

Apertus 8B uses xIELU activations and was trained from scratch with the AdEMAMix optimizer. Post-training combines SFT with QRPO.

Specifications and access

SpecificationValue and source
Weights
Open weightsSource
Parameters
8 billionSource
Context window
65,536 tokensSource
Maximum output
8,192 tokensSource
Knowledge Cutoff
September 2025Source
License
Apache 2.0Source
Inputs
TextSource
Outputs
TextSource
Tool calling
YesSource