Apertus-v1.5-8B-FP8-DYNAMIC

Apertus 1.5

Table of Contents

1. Table of Contents
2. Model Summary
3. Key Features
4. How to Use
5. Evaluation
6. Training
7. Limitations
8. Legal Aspects
9. Contact

Model Summary

Apertus 1.5 is a family of 8B and 70B parameter language models designed to advance the state of multilingual, multimodal, fully open, and transparent AI. The models support a wide range of languages, handle contexts of up to 262,144 tokens, and it uses only fully open training data whilst delivering performance comparable to other models of similar size.

The released models are the result of continued pretraining of Apertus 1.0, adding a multimodal mix of 4T tokens to the 8B model and 2T tokens to the 70B model. Apertus 1.5 thus uses the same architecture as the original release, a decoder-only transformer with the xIELU activation function trained with the AdEMAMix optimizer. Our improved post-training recipe enhances the models' instruction-following and tool-use capabilities and, for the first time, allows developers to enable a thinking mode to improve the models' performance on reasoning tasks.

As a first in the Apertus family, the Apertus 1.5 models support multimodal inputs. The model takes images, audio, and text as input and generates text. This enables many new exciting use cases for our developers.

Key Features

• Fully Open Model: Open weights + open data + open values + full training details including all data and training recipes.
• Massively Multilingual: Supporting a large variety of languages.
• Responsible Development: Apertus is trained while respecting opt-out consent of data owners (even retroactively) where possible and with methods to prevent memorization of training data.
• Native Image & Audio Understanding: Apertus 1.5 introduces multimodal support for processing audio and image inputs, enabling more intuitive and versatile interaction beyond text.
• Reasoning: The models can be switched to thinking mode to reason on the input before generating responses.
• Long Context: Apertus 1.5 by default supports a context length up to 262,144 tokens, a four-fold increase from our initial Apertus 1.0 release.
• Improved Instruction-Following: Significant improvements in instruction adherence ensure more predictable and accurate responses to user prompts.
• Improved Tool Use: Apertus 1.5 has been trained for better tool integration, allowing for more effective use of external tools and APIs.

The technical report with further details along with benchmark results, training pipelines, and intermediate checkpoints will be published in the coming weeks.

How to Use

Users of Apertus can find instructions on getting started with desktop software and cloud providers on our website. Please visit the documentation page if you are interested in trying the model.

Deployment of the models is supported in the following open source frameworks: Transformers, vLLM.

vLLM

We are currently working on adding support for our models to upstream vLLM and transformers releases. In the meantime, you can use our modified version of vLLM and transformers to run the models. We have pre-installed the dependencies in a Docker image which is available in the GitHub Container Registry. The source dockerfile used to build the image is available in this repository. You can pull the image with the following command:

• amd64 architecture:

  docker pull ghcr.io/swiss-ai/vllm_apertus_1.5_release:latest-amd64
  

• arm64 architecture:

  docker pull ghcr.io/swiss-ai/vllm_apertus_1.5_release:latest-arm64
  

You can use the following commands to run the models with vLLM:

• 8B model:

  vllm serve swiss-ai/Apertus-v1.5-8B \
    --chat-template-content-format string \
    --gpu-memory-utilization 0.6 \
    --max-model-len 262144 \
    --enable-auto-tool-choice \
    --tool-call-parser apertus
  

• 70B model:

  vllm serve swiss-ai/Apertus-v1.5-70B \
    --chat-template-content-format string \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.8 \
    --max-model-len 262144 \
    --enable-auto-tool-choice \
    --tool-call-parser apertus
  

Depending on your hardware, you may need to adjust --tensor-parallel-size, --gpu-memory-utilization, and --max-model-len (e.g. lower --max-model-len if you run out of memory). On some hardware configurations, CUDA Graph capture may fail with --tensor-parallel-size > 1 due to the fused all-reduce RMS optimization. If this occurs, launch vLLM with --compilation-config.pass_config.fuse_allreduce_rms false.

<details> <summary>Instructions for Launching Apertus 1.5 with Thinking Mode Enabled</summary>

To enable thinking mode, set --reasoning-parser and --default-chat-template-kwargs.enable_thinking as shown below. The tool-call flags are intentionally omitted: tool calling is unsupported in thinking mode, so we don't recommend combining the two.

• 8B model:

  vllm serve swiss-ai/Apertus-v1.5-8B \
    --served-model-name swiss-ai/Apertus-v1.5-8B-thinking \
    --chat-template-content-format string \
    --gpu-memory-utilization 0.6 \
    --max-model-len 262144 \
    --reasoning-parser apertus \
    --default-chat-template-kwargs.enable_thinking true
  

• 70B model:

  vllm serve swiss-ai/Apertus-v1.5-70B \
    --served-model-name swiss-ai/Apertus-v1.5-70B-thinking \
    --chat-template-content-format string \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.8 \
    --max-model-len 262144 \
    --reasoning-parser apertus \
    --default-chat-template-kwargs.enable_thinking true
  

</details>

Transformers

Apertus 1.5 accepts interleaved text, image, and audio inputs and generates text. The model does not generate audio or images.

The integration is not yet part of a released Transformers version (upstreaming is in progress). Until then, install transformers from our branch:

pip install "transformers[torch,vision,audio] @ git+https://github.com/swiss-ai/transformers.git@3797303dda74844e3d1f8977ff5518bb91f818b4"

Load the processor and model once. The same code works for both released sizes:

import torch
from transformers import AutoModelForMultimodalLM, AutoProcessor

MODEL_ID = "swiss-ai/Apertus-v1.5-8B"  # or "swiss-ai/Apertus-v1.5-70B"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype="auto", device_map="auto"
).eval()


def generate(messages, max_new_tokens=256, **template_kwargs):
    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
        **template_kwargs,
    ).to(model.device)
    with torch.inference_mode():
        output_ids = model.generate(**inputs, max_new_tokens=max_new_tokens)
    return processor.decode(
        output_ids[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True
    )

Text

messages = [
    {"role": "system", "content": "You are a concise and helpful assistant."},
    {"role": "user", "content": "Explain why the sky appears blue in one sentence."},
]

print(generate(messages))

Image

Images can be supplied as URLs, local paths, PIL images, or arrays:

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png",
            },
            {"type": "text", "text": "Describe this image in detail."},
        ],
    }
]

print(generate(messages))

Audio

Audio files referenced by URL or local path are decoded and resampled to 24 kHz automatically; in-memory waveforms are accepted as mono NumPy arrays already sampled at 24 kHz:

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Summarize what is said in this audio clip."},
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
            },
        ],
    }
]

print(generate(messages))

Batching

A batch may mix prompts with different numbers and kinds of media. The shipped tokenizer defaults to the left padding that batched generation requires:

conversations = [
    [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
                },
                {"type": "text", "text": "Describe this image in one sentence."},
            ],
        }
    ],
    [{"role": "user", "content": "Name the four official languages of Switzerland."}],
]

inputs = processor.apply_chat_template(
    conversations,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    processor_kwargs={"padding": True},
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=256)

print(processor.batch_decode(output_ids[:, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Thinking mode

Thinking is turned off by default. Pass enable_thinking=True to apply_chat_template to activate it; the model then reasons between the <|inner_prefix|> and <|inner_suffix|> tokens before its visible answer.

messages = [
    {"role": "user", "content": "A bat and a ball cost 1.10 CHF together. The bat costs 1 CHF more than the ball. What does the ball cost?"},
]

print(generate(messages, enable_thinking=True, max_new_tokens=2048))

Budget generously for max_new_tokens in thinking mode: the reasoning can be several times longer than the visible answer, and a truncated generation may end before the answer begins. The reasoning markers are deliberately not stripped by skip_special_tokens=True, so the deliberation span can be parsed out of the decoded text.

Input notes

• Each image or audio content block corresponds to exactly one media item; placeholder/media count mismatches raise an error instead of being silently reassigned.
• Audio contributes 40 tokens per second.
• The LM head covers the 131,072 text tokens (output_vocab_size); logits are padded to the full 266,752-token vocabulary with non-selectable scores, so standard generation utilities work unchanged while image and audio tokens are never generated.
• The vision and audio tokenizers are precision-sensitive and stay in float32 automatically on half-precision loads. Avoid re-casting the loaded model with .half()/.to(dtype).
• Native video inputs are not supported; applications may extract frames and pass them as images.

Evaluation

Detailed benchmark evaluations for the pretraining and post-training phases, multilingual evaluations, and long-context evaluations from the original Apertus release will be provided in the technical report.

Text benchmarks cover various domains like knowledge (MMLU, MMLU-Pro, AGIEval, ARC-Challenge, TruthfulQA, CommonsenseQA, SQuAD v2, HellaSwag), instruction following (IFEval, Multi-IF, IFBench), chat (AlpacaEval), math (GSM8K, MATH, Minerva-MATH, MathQA, GSM8K-Platinum, MATH-500), coding (HumanEval, MBPP), reasoning (BIG-Bench Hard, GPQA, ACPBench, DROP), multilingual ability (Global-MMLU, MGSM, multilingual TruthfulQA, INCLUDE, mLogiQA, XNLI, multilingual ARC), cultural knowledge (BLEnD, CulturalBench, SwitzerlandQA), safety and bias (BBQ, ToxiGen, WMDP, MultiJail, PolygloToxicityPrompts), system-prompt robustness (RealGuardrails), and tool use (BFCL v3).

Image benchmarks cover general VQA and perception (GQA, MMStar, RealWorldQA, VQAv2, V*Bench, CountBench), robustness and bias (MMVP, VLMs-Are-Biased, VLMsAreBlind), documents and OCR (ChartQA, DocVQA, SEED-Bench-2-Plus, InfoVQA, OmniDocBench), visual math and logic (BabyVision, MathVision, MathVista), STEM knowledge (AI2D, MMMU, MMMU-Pro, ScienceQA), spatial intelligence (MMSI, ViewSpatial, MindCube, EmbSpatial), remote sensing (FRIEDA, GeoBench, VRSBench-VQA), hallucination alignment (POPE), and medical imaging (PathVQA, PMC-VQA, SLAKE, VQA-RAD).

Training

Model

• Architecture: Transformer decoder
• Pretraining tokens: 17T for the 70B model, 19T for the 8B model
• Precision: bfloat16

Limitations

Apertus can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.

See the Apertus Charter for the alignment principles that govern the responses of the models.

Legal Aspects

EU AI Act Transparency Documentation and Code of Practice

Documentation and code of practice for the EU AI Act are available at the following links:

• Apertus_1_5_EU_Public_Summary.pdf
• Apertus_1_5_EU_Code_of_Practice.pdf

Data Protection and Copyright Requests

For removal requests of personally identifiable information (PII) or of copyrighted content, please contact the respective dataset owners or us directly:

• [email protected]
• [email protected]

Currently no output filter is provided. We endeavour to address data protection deletion requests by updating filters in our data pipeline, and providing regular (at least bi-annual) releases to the model weights. We strongly advise downloading and updating new model weights from this site every six months.

Contact

To contact us, please use the links at https://apertus-ai.org/contact/ or send an email to [email protected].

FP8 quantization

This checkpoint was quantized to FP8 with llm-compressor by Artificialy SA. We evaluated two schemes — FP8-DYNAMIC and FP8-BLOCK — against the BF16 reference across eight text benchmarks.

TL;DR: FP8-DYNAMIC recovers 99.1% of the BF16 macro score with throughput at or above BF16 parity on most workloads; FP8-BLOCK recovers 97.8% and is slower than BF16 on most workloads under current Hopper kernel support. FP8-DYNAMIC is the recommended checkpoint.

Evaluation setup

• Reference: the BF16 Apertus-v1.5-8B checkpoint. Both FP8 schemes are calibration-free (data-free).
• Framework: NeMo Evaluator v0.3.0 over a vLLM server; up to 500 problems per benchmark, temperature=0.8, top_p=0.9, n_repeats=1, 32 concurrent requests.
• Hardware: 2× NVIDIA H100 NVL (tensor-parallel 2). The reference and both FP8 variants used identical server flags; only the checkpoint changed.

Quality

Per-benchmark score, with recovery vs. the BF16 reference in parentheses:

Quantized-vs-reference deltas fall within each benchmark's confidence interval; small positive deltas reflect evaluation variance, not a real gain from quantization. recovery = candidate / reference × 100 is a descriptive ratio, not an equivalence test.

Throughput

Generated tokens/s relative to the BF16 reference, under identical server flags:

FP8-BLOCK is slower than BF16 on most workloads — this reflects current Hopper kernel support, not block quantization in general (the scheme is Blackwell-optimized). MGSM and DROP are omitted above: both showed implausibly large speedups (up to ~9×) that indicate anomalously slow reference runs and need separate profiling.

Recommendation

FP8-DYNAMIC gives the best quality–throughput trade-off: 99.1% macro-score recovery, a lower aggregate failure rate than the reference, better quality retention than FP8-BLOCK on five of eight benchmarks (largest gaps on MATH-500 and GSM8K), and throughput at or above BF16 parity on six of eight workloads.