Apertus 1.5 adds image/audio understanding, longer context and a thinking mode to the open 8B/70B models
In short: Swiss AI released Apertus 1.5, a continued pretraining of its open Apertus 1.0 models (8B and 70B, Apache 2.0) with native image and experimental audio understanding, a 4x larger context window of 262,144 tokens, an optional thinking mode, and improved instruction-following and tool use. The models were trained on a total of 19T (8B) and 17T (70B) tokens, with full training data, recipes and intermediate checkpoints published for reproducibility. The team also published detailed, standardized evaluation cards and announced early API access from providers like Swisscom and Infomaniak, plus a roadmap for a ~600B-parameter MoE Apertus 2.0.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Adds native image and audio input (audio support is experimental); the models still generate only text
- 2Context window grows 4x to 262,144 tokens
- 3New optional 'thinking mode' for step-by-step reasoning before answering
- 4Improved instruction-following and tool use, trained on execution-verified tool trajectories
- 5Total training data grows to 19T tokens (8B model) and 17T tokens (70B model)
- 6Full training data, intermediate checkpoints and standardized Eval Cards released publicly
- 7Early API access announced via Swisscom, Infomaniak, Phoeniqs and others; apps like Proton Lumo and Infomaniak Euria already integrate it
- 8Apertus 2.0 roadmap announced: sparse/MoE architecture, ~600B parameters, focus on agents, math and coding
| Parameter | Before | Now |
|---|---|---|
| Context window | Not specified (Apertus 1.0 baseline) | 262,144 tokens (4x increase over Apertus 1.0) |
| Input modalities | Text only | Text, images, and audio (audio understanding experimental); generation remains text-only |
| Reasoning | No explicit reasoning mode | Optional 'thinking mode' that reasons before answering, boosting performance on tasks like AIME |
| Tool calling | Standard tool use | Improved, trained on execution-verified tool-use trajectories plus RL/DPO tuning |
| Training data volume | Apertus 1.0 pretraining corpus | 8B model: 19T total tokens (+4T); 70B model: 17T total tokens (+2T) |
| License | Apache 2.0 | Apache 2.0 (unchanged, fully open weights/data) |
Why it matters
For teams building agents or voice/contact-center automation, a fully open, Apache 2.0-licensed model with multimodal input, long context, and verified tool-use training is a meaningful alternative to closed APIs — especially where self-hosting, cost control, or data residency matter. The experimental audio understanding and improved tool calling are directly relevant to pipelines that combine speech, documents and external function calls.
What it means for AI agents and contact centers
This is worth evaluating as a self-hostable backbone for agent or tool-calling workflows where you need full control over data and infrastructure instead of a closed API — test its tool-calling reliability, long-context handling of call transcripts or documents, and the experimental audio understanding against your current STT/LLM stack, keeping in mind it still trails frontier closed models on raw benchmarks.
🧪 Worth Testing
Open licensing, long context, improved tool calling and experimental audio input make it a candidate for self-hosted agent and voice pipelines, provided its multilingual and audio capabilities hold up in direct testing against your current tools.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Apertus 1.5 Release Notes“15 Oct 2026, 12:08
- Published by source
- 15 Oct 2026, 12:08
- Found by our system
- 15 Oct 2026, 16:41
- Summary generated
- 15 Oct 2026, 16:43
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy