Aleph Alpha's open Kolibri uses 11.2% fewer German tokens than GPT-5's tokenizer

Top News · 410 HN points

Need a strong German and English model you can run on your own hardware? Aleph Alpha launched Kolibri 1 as an open-weight model trained from scratch in Germany and Finland. In Aleph Alpha's evaluation, it scored 75.5 in English and 70.8 in German overall, ahead of Qwen3.5 35B-A3B at 74.7 and 69.8.

Here's what changed:

• Its UniBPE tokenizer uses 11.2% fewer tokens on German text than GPT-5's tokenizer, in Aleph Alpha's measurement.
• On AIME 2025 it hit 96.9 in English and 87.5 in German, against 89.6 and 84.4 for the closest model with about 3B active parameters.
• It is an MoE with 78.1B total and 3.46B active parameters per token.
• Context is 262,144 tokens natively, validated up to 1,048,576 tokens.
• On Omniscience, it admitted uncertainty or gave a partial answer 44% of the time.

One catch: coding and multi-turn tool calling trail rivals. It scored 66.4 on SWE-bench Verified versus 73.8 for Qwen3.6 35B-A3B.

Try it: pip install 'aleph-alpha-inference>=1', then vllm serve Aleph-Alpha/Kolibri-1 (Apache 2.0 weights on Hugging Face, about 78 GB of GPU memory).

Sources: tej.as · tej.as