Tasks and inputs
Apertus 1.5 8B is listed for Audio-text-to-text, Image-text-to-text, Multimodal, Text generation. It accepts text, image, audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Apertus 1.5 8B is the Swiss AI Initiative's open multilingual model, now taking text, images and audio with a 262K-token context window.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No first-party price. Registry lists a stepping-stone route at about $0.024 in / $0.119 out, unverified.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
The original Apertus 8B was text-only. Apertus 1.5 8B is not. The Swiss AI Initiative continued pretraining on about 4 trillion multimodal tokens, bringing the total to 19 trillion, and the model now accepts text, images and audio while still answering in text.
It also gains reach. The default context is 262,144 tokens, four times the 1.0 release. For a model this small, that is generous enough to hold long contracts or transcripts without chunking.
Some coverage describes audio support as experimental. Test it on your recordings before relying on it.
Thinking mode is off by default. Turn it on with enable_thinking=True in the chat template, and the model writes its reasoning between dedicated inner tokens before answering. The reasoning can run several times longer than the answer, so give it a generous token budget.
The catch: tool calling is not supported in thinking mode. Pick one per request. For agent loops that call tools, keep thinking off.
| Item | Detail |
|---|---|
| Hugging Face | swiss-ai/Apertus-v1.5-8B |
| Size | 8B (Hub metadata shows ~9B) |
| Context | 262,144 tokens |
| Licence | Apache-2.0 plus acceptable use policy |
| Access | Gated; accept the policy to download |
| vLLM tool flags | --enable-auto-tool-choice --tool-call-parser apertus |
Upstream libraries are not there yet. Serve it with the project's modified vLLM build or its prebuilt Docker image, or install Transformers from the project branch. Quantised GGUF builds exist for local tools. No hosted inference provider serves it at the time of writing.
On the published card, it scores 78.47 on GSM8K and 48.75 on MMLU-Pro. Those are the developers' figures.
It is unclear. Some coverage says 24 July 2026, while the card shows a paper on 11 October and release notes on 15 October. We use the October date.
Apertus is built around data compliance: training honoured opt-outs where possible, and the team suggests refreshing weights about every six months as removals are processed. If provenance matters to your legal team, that is a real differentiator. Test it against your languages, not just English, and compare it with its larger sibling Apertus 70B or other multimodal models. See every release on the Swiss AI provider page.
Compare the model on the work you need to complete.
Apertus 1.5 8B is listed for Audio-text-to-text, Image-text-to-text, Multimodal, Text generation. It accepts text, image, audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0 plus acceptable use policy (gated). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
Alpamayo 2 Edge is NVIDIA's open 2B-class driving VLA that outputs trajectories with causal explanations, sized for DRIVE AGX Thor in-vehicle use.
shanghai-ai-lab
Sage-SciPerception is Shanghai AI Lab's set of projectors feeding crystals, spectra, biological sequences and biosignals into a language model's input.
inclusionAI
Ling 3.2 Flash is inclusionAI's 560B-A25B multimodal MoE with a 1M-token window, built for visual agents, 3D tasks and medical image reasoning.
kyutai
REKI-2B is Kyutai's open 2.4B robot vision-language-action model with cheap register-token memory, pretrained only on public data.
black-forest-labs
FLUX 3 Action is Black Forest Labs' open 7B world action model for robots, predicting action chunks and future video from cameras, state and text.