Swiss AI Details Apertus 1.5: 8B and 70B Multimodal Models With 262K Context

Swiss AI published the Apertus 1.5 evidence package on October 15

The Swiss AI Initiative—an ETH Zurich and EPFL collaboration supported by the Swiss National Supercomputing Centre—published detailed Apertus 1.5 release notes on October 15, 2026. The release covers two final model IDs: swiss-ai/Apertus-v1.5-8B and swiss-ai/Apertus-v1.5-70B. The companion technical report was submitted to arXiv on October 11.

There is an important chronology caveat. The release notes say updated weights were released on July 24, and the Hugging Face API records repository creation on that date. The current records show both final repositories modified on October 14. Today’s event is the public release-note and evidence package, not the first upload of the weights.

What is available now—and what access still requires

The two final Hugging Face repositories are live at swiss-ai/Apertus-v1.5-8B and swiss-ai/Apertus-v1.5-70B. At review time, their API records identified 8.9 billion and 72.0 billion total parameters respectively, tagged both as image-text-to-text and Apache-2.0, and exposed model-weight, image-tokenizer and audio-tokenizer files. The pinned commits reviewed were ef4bbc54d6fe658bc782280ff9f3b53a10a9758f for 8B and 94888dd0db53f674b4c5304c1b726a7783d21393 for 70B.

“Open weights” does not mean anonymous one-click retrieval here. Both repositories were auto-gated. Hugging Face’s live card metadata required an authenticated user to agree to the Apache 2.0 license, the Apertus acceptable-use policy and a privacy policy, and to share contact details with the repository authors. The release notes also point to hosted integrations and announced developer APIs from providers including Swisscom and Infomaniak, but they do not give one universal endpoint, price, quota or service-level contract. Buyers must inspect the selected provider’s current offer.

Images and audio enter the same token sequence as text

Apertus 1.5 adds image and audio understanding while retaining text-only generation. The project uses early fusion: an Emu3.5 tokenizer turns a 512-by-512 image into 1,024 discrete tokens, and WavTokenizer represents audio at 40 tokens per second. Those tokens enter the language model’s sequence with text; the architecture does not add a separate vision encoder.

This design makes multimodal input part of the context budget. A ten-second audio clip represents roughly 400 audio tokens before prompt text or generated output, while a 512-by-512 image represents 1,024 tokens under the documented setting. Those conversion rates help estimate sequence occupancy, not end-to-end memory, latency, accuracy or cost. The project explicitly calls spoken-language support experimental and the models do not generate images or audio.

The context window is four times larger, but length is not recall

The documented maximum context grows from 65,536 to 262,144 tokens. The project says it extended training progressively and then ran long-context supervised fine-tuning because short-context post-training can erode long-context behavior. Optional thinking is selected through the chat template rather than a separate checkpoint. The release also reports execution-verified tool-use trajectories during supervised fine-tuning.

A context-window specification is a capacity ceiling, not proof that the model can find, combine and faithfully cite evidence throughout that window. Teams should test position-dependent retrieval, conflicting evidence, multi-document synthesis, tool-schema adherence, refusal behavior and total accelerator memory at the exact sequence lengths they intend to use. Thinking tokens, media tokens and output all compete for operational resources even when the advertised context accepts the request.

Training artifacts are unusually broad, while results remain project-reported

The project describes continued pretraining with four trillion additional tokens for the 8B model and two trillion for 70B, bringing stated totals to 19 trillion and 17 trillion. Post-training proceeds through supervised fine-tuning, long-context SFT, reinforcement learning with verifiable rewards, preference optimization and, for the 8B final model, charter alignment. Swiss AI links base, SFT, RLVR and DPO checkpoints, plus data mixes, preference data, recipes and evaluation code.

That artifact trail is useful for audit and reproduction, but it does not convert the reported scores into independent results. The paper reports gains in instruction following, reasoning, multilingual work, document and spatial understanding. It also says frontier open-weight models remain ahead overall and specialized models remain stronger on perception and high-resource speech transcription. AccessAllGPT did not rerun the evaluation cards or determine whether all compared systems used equivalent prompts, compute and inference settings.

Pricing depends on the deployment path

Downloading gated weights does not carry a per-token model fee from the repository, but self-hosting still incurs accelerator, storage, bandwidth, engineering and operations costs. The 70B repository’s parameter count alone signals a materially different memory and serving problem from 8B; actual requirements depend on precision, quantization, parallelism, context, concurrency and the multimodal tokenizers.

The release notes name hosted applications and companies that have announced API access, but publish no single Apertus API identifier or price table that applies across them. Do not copy a provider’s price into an “Apertus price” field without binding it to provider, region, model variant, processing tier, input modality, context and date. Compare completed accepted tasks, including media preprocessing, long-context prefill, retries and tool calls.

What builders should do next

Trial the 8B model first when local control, multilingual coverage or transparent training artifacts matter and your workload can tolerate a self-managed stack. Freeze the repository commit, license and policy snapshots, runtime, tokenizer and chat template. Build separate text, image, audio, long-context and tool-use case sets; record per-case quality, unsafe behavior, latency, peak memory, tokens, failures and complete cost.

Move to 70B only if a controlled comparison shows an outcome gain worth the operating cost. Constrain audio to experiments until its quality clears task-specific gates. Wait for a hosted provider contract when your team cannot own inference operations. Reject deployment where repository terms, data handling, required hardware, text-only output or measured reliability conflict with the product contract.

Disclosures

AccessAllGPT received no model access, weights, API credits, hardware, data, benchmark logs, briefing, review, payment or compensation for this article. The Swiss AI Initiative, ETH Zurich, EPFL, CSCS, Hugging Face and named providers did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Apertus 1.5. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with the organizations cited. Publication-wide relationships are listed on the disclosures page.