Future AI Pulse
Quick answer: Swiss AI’s new Apertus 1.5 release notes explain multimodal inputs, optional thinking and more detailed evaluation reporting. The notes were published October 15, 2026; they say the 8B and 70B weights were released July 24. This is a documentation update about existing weights, not a claim that those weights first launched today.
Capabilities and boundaries
The official release notes describe image and audio understanding alongside text, a 262,144-token context window and optional thinking mode. The models generate text only. Speech processing remains experimental. The project says the weights use Apache 2.0 and links model cards, training materials and intermediate checkpoints.
Swiss AI reports improvements over Apertus 1.0 while acknowledging that the release still trails frontier open-weight models overall. A developer’s reported benchmark result does not establish performance on every document, language or tool workflow. Future AI Pulse has not run these models.
Input support does not imply output support
A system may accept a photograph or recording and respond with text. That is different from generating a new image or speaking an answer. Keep those directions separate when choosing a model for an application.
For a document task, write down whether you need to read scanned text, understand a chart or describe a photograph. Those are different requirements even though they all involve images. Likewise, recognizing words in a recording and interpreting a spoken request should not be treated as one interchangeable capability.
Long context needs a task-level check
A larger window describes how much input a system can accommodate under its supported setup. It does not guarantee that every relevant detail will be used correctly. Test whether an answer identifies the right passage, preserves qualifications and admits when the document does not contain the answer.
Start with a document whose important facts you know. Place the required evidence in different sections and include a question with no supported answer. Check source locations manually before extending the workflow to consequential material.
Read evaluation records, not just rankings
The release notes describe machine-readable evaluation records that include the evaluator, access method, generation settings and metric interpretation. Those details help explain why apparently similar scores can differ. Follow the linked records before comparing a headline number with another model’s result.
For your own comparison, preserve the checkpoint, runtime, inputs and settings. Distinguish a result you measured from one reported by the project. If thinking mode changes the setup, record it rather than presenting both modes as the same run.
Plan a small comparison for this exact task: [task]. Record the checkpoint, runtime and thinking-mode setting. Use the same inputs and acceptance criteria for each run. Check image or audio handling separately from text-only behavior. For document answers, retain source locations and unsupported questions. Separate developer-reported scores from measurements we actually produce. Do not claim model execution without actual run output.
Choose the artifact that matches the job
A base checkpoint, an intermediate training checkpoint and a final assistant checkpoint have different purposes. Read the relevant model card and quickstart before assuming that any file in the collection is a ready-to-use chat assistant. Confirm the supported runtime and resource requirements for the specific artifact.
Openness makes inspection and reproduction possible; it does not remove the work of checking whether a deployment fits your task. The checklist above is an original planning aid, not a reproduced Apertus benchmark.
Related reading: our training-brief guide covers baseline evidence, and the local-agent guide discusses runtime and workflow checks.
Illustrative server photograph: Kevin Ache / Unsplash, under the Unsplash License. Not Swiss AI infrastructure.
Sources
Source checked October 15, 2026. Written with AI assistance. No model execution or independent benchmark reproduction is claimed.