Releasing eustella version 1.5 - insights and thoughts from the eustella Team

This is the last of three research articles we are releasing in connection with eustella version 1.5. This third piece acts as the summary of, and the why behind, the two more in-depth articles we published. The reported evaluations here were already part of our internal model-evaluation process; we publish them now because we believe that a sovereign AI stack is only as trustworthy as its evaluations are open. We are publishing the method and the numbers so that the results can be examined, evaluated, and criticised. More, and deeper, evaluations will follow.

You can find all the benchmarks, sources and references cited at the end of this article.

For more in-depth comparisons and methodology, take a look at Measuring the political bias of eustella v1.5 against its base model Qwen3.8-27B and In-Depth Comparison and Benchmarks of eustella v1.5.

Introduction

eustella v1.5 is our new AI System we developed for our Agent Platform eustella. Our new version includes an optimized harness, weight quantizations for near-zero quality loss as well as RL and other post-training steps. Our optimizations can be applied to almost all open-weight models. With this new release we want to take our next step as a European AI company. First we are trying to solve the infrastructure supply chain problem and now we will turn more and more to post-training and optimizing open-weight and open-source models until the market and technology allow for completely pre-trained models made by eustella. This third article about eustella v1.5 shall mark the beginning of this journey, with more in-depth innovation and research already in progress.

A quick Summary of the results

eustella v1.5 shows extremely promising results, both in the performance on the selected benchmarks we use for our internal evaluations and in the first proper look into the political / ideological bias of open-weight models.

On the performance side, eustella v1.5 leads its base model Qwen3.8-27B on all 14 benchmarks we ran for both models under identical settings, on average by 5.3 points. The biggest jumps are in programming (LiveCodeBench v6, from 72.0 to 89.1), in following complex instructions (IFBench, from 71.67 to 82.6) and in complex multi-turn conversations (MultiChallenge, from 63.55 to 74.3).

On the political side, the base model presents the Chinese government's narrative as fact in 30.4% of its answers on topics the Chinese government censors. With eustella v1.5 that share falls to 1.9% - roughly 16 times less often. Asked to criticise the Chinese government, the base model fully carries out the request in only 48.9% of cases, while eustella v1.5 does so in 84.4% of cases, much closer to how it treats the governments of the US, Russia or the EU. At the same time, eustella v1.5 did not become one-sided in the other direction: on Anthropic's even-handedness benchmark for US politics, both models are at the same level (92.3% vs 91.7%, no significant difference).

All numbers, their sources and the limitations of both evaluations are explained in detail in the two other articles.

Our thoughts on the current State of European AI

Europe is making an effort. More and more startups sprout, research is still Europe's strongest root and compute capacity is being built. Aleph Alpha released its newest model, which gained traction after a long time, and Mistral just launched its newest frontier model with Mistral Large 4. The Open Source initiative of the German Soofi model series tries to push the mission of fully open source models. Europe is warming up to the AI race more and more. Are the current efforts going to be enough? Not yet - but that is a different question entirely right now, because Europe needs to show what it is capable of in this day and age, even if it may still lag behind the global frontier of the current transformer architecture. The important part is that we continue to push and stretch the boundaries of what we as a united Europe are capable of to secure a safe future in the age of AI. We at eustella truly believe that just having access to SOTA intelligence is not enough. We must build on top of the resources we currently have and continue forward to expand our own capabilities. That is why we want to set our own first stepping stone with eustella v1.5 to continue on our own way of what AI may mean for us and the world as a whole. One step at a time.

The release of eustella v1.5

We built eustella v1.5 with two key principles as the driving decision makers: 1. increase the performance compared to the open-weight models we currently have access to and 2. do that with the limited resources we have available as a small team here in Vienna, Austria. That is why we based eustella v1.5 mainly on (Chinese) open-weight models, specifically the dense 27B Parameter Model Qwen3.8-27B, showcasing the performance that can be squeezed out of such a comparatively small model while trying to even out points of criticism of using Chinese open-weight models. We took general safety alignments very seriously from the beginning of eustella inside its - what is nowadays called - harness. However, now we can also show with the numbers outlined above that - with both principles in mind - European AI can work on top of open-weight models, also from China.

Regarding the infrastructure stack: We worked with only one GPU Provider for the development of eustella v1.5, namely Verda based in Finland (which just recently closed its Series B!) - keeping the whole process and most of the value chain here in Europe (apart from the obvious GPUs themselves, which are NVIDIA Chips).

Incremental Steps

Just like we built eustella - as a European alternative because we ourselves did not see any alternative but to do so - we built the new version of eustella as the next logical step. In the future we want to optimize more and more LLMs while also diving deeper into other systems for voice, transcriptions, images as well as so-called system 1 intelligence models (or better said: sophisticated classifiers).

Further Information

For the in-depth comparisons of eustella please take a look at the second piece of this article series here and for the detailed methodology of our alignment evaluation and how we approach this and similar topics please take a look at this article.

References

Our other articles

• Rosoli, M., & Maitz, A. (2026, 20 October). Measuring the political bias of eustella v1.5 against its base model Qwen3.8-27B. eustella Research. eustella.com/research/measuring-political-bias-of-qwen3-8-27b
• Rosoli, M., & Maitz, A. (2026, 20 October). In-Depth Comparison and Benchmarks of eustella v1.5. eustella Research. eustella.com/research/eustella-v1-5-full-comparison

Short list of References for the three articles

• Anthropic (2025, 13 November). Measuring political bias in Claude. https://www.anthropic.com/news/political-even-handedness
• Jain, N., et al. (2024). LiveCodeBench: Holistic and contamination free evaluation of large language models for code. arXiv:2403.07974
• promptfoo (2025, 28 January). 1,156 Questions Censored by DeepSeek. https://www.promptfoo.dev/blog/deepseek-censorship/
• Pyatkin, V., et al. (2025). Generalizing verifiable instruction following (IFBench). arXiv:2507.02833
• Qwen Team, Alibaba Cloud (2026). Qwen3.8-27B. https://huggingface.co/Qwen/Qwen3.8-27B
• Sirdeshmukh, V., et al. (2025). MultiChallenge: A realistic multi-turn conversation evaluation benchmark challenging to frontier LLMs. arXiv:2501.17399
• SpeechMap.ai (xlr8harder) (n.d.). https://speechmap.ai

The full benchmark list, every published score with its source, and the complete references of both studies are in the two articles above.