Hugging Face releases Apertus 1.5 with multimodal, long-context, and reasoning upgrades

Hugging Face published release notes for Apertus 1.5, extending the 8B and 70B open-weight models with native image and audio understanding, an optional "thinking mode" for reasoning, a 262,144-token context window (4x longer than v1.0), improved instruction-following, and better tool use.

The updated 8B and 70B weights were released to Hugging Face on July 24, 2026 under an Apache 2.0 license, with further details available in a technical report. Apertus 1.5 continues pretraining by adding 4 trillion tokens to the 8B model and 2 trillion tokens to the 70B model, reaching 19T and 17T tokens respectively.

Multimodality is implemented via an early-fusion tokenizer that treats images and audio as discrete tokens interleaved with text, using the Emu3.5 tokenizer for images and WavTokenizer for audio. The architecture remains unchanged, enabling text-only-speed training while supporting image and audio inputs.

Post-training consists of five stages: supervised fine-tuning on 3.7M text samples plus image and speech data, long-context SFT at 262k tokens, reinforcement learning with verifiable rewards on ~350k prompts, DPO for chat quality and tool use, and SDPO to align the model with the Apertus Charter.

Evaluation was expanded to include multimodal benchmarks (via MLLM-eval-suite), text-based tasks, translation and multilingual tests, safety, and tool-use metrics. Apertus 1.5 8B outperforms similarly sized open models on instruction-following and multilingual tasks, and thinking mode improves performance on reasoning-heavy benchmarks like AIME.

The team reports multimodal performance competitive with open-weight peers such as Gemma 4 12B, with strong spatial and document/chart understanding. On the Last Translation Benchmark, the 8B variant matches GPT-OSS-20B and surpasses Claude Haiku 4.5 and Google Translate.

Apertus 1.5 publishes evaluation results through the EvalEval Coalition using the Every Eval Ever (EEE) schema, which records evaluation settings, access methods, and metric interpretations to ensure reproducibility. The EEE schema aligns with governance frameworks like the EU AI Act and the GPAI Code of Practice.