DE/ENCommissioned Battle

Neutral LLM benchmark

Six models.One topic.Judged blind.

KI-Battle pits language models against each other on a single topic — they grade one another blind and actively hunt for mistakes. Not a vibe ranking, but a data-driven third-party benchmark.

Latest battlesRequest a battle
Live ranking · sample runDeutsche Rechtschreibung und Grammatik →
  1. Claude Opus 4.8☁︎ Cloud
    0,782
  2. GPT-5.5☁︎ Cloud
    0,767
  3. GLM-5.2🖥︎ Local
    0,765
  4. GPT-OSS 120B🖥︎ Local
    0,671
  5. Gemma 4 31B🖥︎ Local
    0,616
  6. Gemma 3 27B🖥︎ Local
    0,423

Why KI-Battle is different

🎭

Judged blind

Models see anonymised answers and rank each other — without knowing who is who. Peer review instead of gut feeling.

☁︎

Cloud vs. local

Cloud heavyweights face locally-run models. We show the VRAM footprint — down to the question: what runs on 16 GB at home?

🔍

Full transparency

Every full answer, every challenge, every score — open to inspect. The video shows the highlights, the site the complete run.

From the blog

Bilingual summaries of every battle.

See all articles
·🇩🇪 DE

Deutsche Rechtschreibung und Grammatik

Winner: Claude Opus 4.8

Six language models competed on German spelling and grammar — cloud heavyweights against locally-run models. Claude Opus 4.8 edged out GPT-5.5 by a razor-thin 0.015 points, decided by a higher hit rate in the challenge phase. The real surprise was the locally-run GLM-5.2, which reached the podium and left cloud-versus-local almost tied. The ≤16 GB home category stayed empty: even the smallest model tested needed around 17 GB. GPT-5.5 actually led on peer recognition and consistency — the challenges made the difference.

View battle →
·🇩🇪 DE

Logisches Denken & Schlussfolgern

Winner: GPT-5.5

The second battle tackled formal logic, lateral thinking and self-reference. GPT-5.5 narrowly beat Claude Opus 4.8 — but the real story was the home category: for the first time the ≤ 16 GB class had a winner, with Phi-4 14B running on a regular consumer GPU at around 11 GB. On the easier logic questions the field stayed tight; only the barber paradox separated the top from the rest. (Placeholder entry with illustrative values.)

View battle →

The method

Dialectic Convergence

Five phases: questions the models write themselves, blind grading, targeted error-hunting and a consistency check. That is how opinion becomes a defensible ranking.

Understand the method
  1. 1Questions
  2. 2Answers
  3. 3Blind ranking
  4. 4Challenges
  5. 5Consistency

For companies

A neutral benchmark on your task

Your task, your models, our independent procedure. As a published battle or strictly confidential under NDA.

Request a battle