One-Line Summary

Open, multimodal models move into practical deployment while agent skills and a new finance benchmark push evaluation beyond plain text.

LLM & SOTA Models

Apertus 1.5 adds image and audio, long context, and 'thinking' mode

Apertus 1.5 is a family of 8 billion and 70 billion parameter large language models (LLMs) that now accept images and audio alongside text and can keep track of up to 262,144 tokens in a conversation. The team continued pretraining with 4 trillion tokens for the 8B model and 2 trillion tokens for the 70B model, enhanced instruction-following and tool use, and introduced an optional “thinking mode” that lets the model deliberate before answering to improve reasoning. The project emphasizes full openness: weights, data sources, and training details are disclosed. ¹

Weights are released under the Apache 2.0 license, but access is gated: users must accept an Acceptable Use Policy (AUP) and agree to share contact information (email and username) with the authors; a published privacy policy covers this data. This strikes a balance between open weights and responsible use tracking. ¹

You can run Apertus via the vLLM server (an open-source inference engine) or a custom Transformers branch; upstream integration is in progress. The docs provide Docker images for AMD64 and ARM64, note that “thinking mode” disables tool calling, advise budgeting more output tokens for reasoning, and explain practical limits (for example, audio counts as 40 tokens per second and you may need to adjust graphics processing unit (GPU) memory settings). ¹

The release sits within a broader Swiss “sovereign AI” effort led by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). Reported use cases span government translation services, medical model building (MeditronFO), and a newsroom workflow at Bajour; CSCS is also launching an inference application programming interface (API) for the Swiss academic community. ²

Open Source & Repos

K-Dense updates BYOK and Scientific Agent Skills for 'AI scientist' workflows

Scientific Agent Skills is a library that adds ready-made scientific capabilities to agents: it lists 150 skills, taps into 100+ scientific databases, and claims usage by 160,000+ scientists. The v2.55.0 update refreshes documentation and formatting, redesigns scientific writing safeguards, overhauls PPTX poster generation, and notes compatibility with Cursor, Claude Code, Codex, and Google Antigravity as well as the open Agent Skills standard. ³

K-Dense BYOK is a desktop “AI co-scientist” app that shipped v0.9.0 on 2026-07-24 under the MIT license; the repo highlights 149 skills, 326 workflows, and 229 databases integrated into the environment. For research teams, these packaged skills and workflows can reduce setup time for literature reviews, data queries, and figure/poster drafting. ⁴

Research Papers

FinanceComplexQA: Agent reasoning benchmark for real-world financial docs

FinanceComplexQA is a new benchmark that tests how well AI agents handle industrial-grade financial documents. The authors first build Finance-LaTeX SKILL to synthesize 2,000 professional documents and 6,000 question–answer pairs, then assemble 2,026 deep-research tasks targeting 1,009 documents, with bilingual support, six scenarios, and seven tasks that require expert-level document reasoning. ⁵

Using FinanceComplexQA, they evaluate leading retrieval-augmented generation (RAG) systems and agentic reasoning tools for document QA with an Agent-as-a-Judge setup and multiple metrics. The study surfaces failure modes in numerical computation, multi-hop reasoning, content summarization, and industry analysis—pinpointing where agents still struggle on complex, messy inputs. ⁵

Why It Matters

Apertus 1.5 packages long-context, multimodal understanding with transparent training and a compliance gate, signaling a path for public institutions and enterprises that need both capability and auditability. The Swiss program and CSCS-hosted access also show how national infrastructure can distribute open-weight models at scale. ²

At the same time, domain skills libraries and the FinanceComplexQA benchmark push agent evaluation beyond plain text and into realistic scientific and financial workflows—helping teams choose tools that actually hold up on multi-document, calculation-heavy tasks. ⁵