Apertus 1.5 Release Notes
- ID
- 35816
- Status
- summarized
- Published
- 15 Oct 2026, 5:08 PM
- Fetched
- 15 Oct 2026, 10:26 PM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/swiss-ai/apertus-v15-release-notes
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 7.0
- Created
- 15 Oct 2026, 10:26 PM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_startup_founders
What happened
Apertus 1.5 extends the 8B and 70B Apertus 1.0 models with image and audio understanding, an optional thinking mode, a 262,144-token context window (4x longer), improved instruction-following, and better tool use. The updated weights were released on July 24, 2026 under Apache 2.0, after continued pretraining added 4T tokens to the 8B model and 2T to the 70B model, for totals of 19T and 17T. Multimodality uses an early-fusion tokenizer approach—Emu3.5 turns a 512×512 image into 1,024 tokens and WavTokenizer encodes audio at 40 tokens/sec—so the models understand images/audio but generate text only.
Why it matters
If you build agents or document, diagram, or photo workflows, you can test self-hosted multimodal understanding with Apache 2.0 8B/70B weights and a 262k context, but you should not plan on audio generation or production audio pipelines because spoken-language support is experimental and output is text-only. The open weights and training details also make it worth comparing against API-only multimodal models for cost, data control, and tool-calling workflows.
Discussion angle
Where would Apertus 1.5's early-fusion tokenizer approach (images/audio as tokens, text-only output) actually beat a separate vision encoder or a hosted multimodal API in an agent pipeline, and where does the lack of audio generation make it a non-starter?