All Benchmarks/Safety & alignment

Nejumi 4 — ALT (alignment)

JapaneseHigher is better

Nejumi 4's alignment (ALT) average: an equal-weight average of six areas in Japanese, namely following instructions, ethics, toxicity, bias, truthfulness, and robustness. It covers instruction-following and robustness, not just safety. Scores run from 0 to 100%. Higher is better.

Other versions:Japanese MT-Benchjaster (0-shot)Nejumi 4 — GLP (general language)Nejumi LLM Leaderboard 4
Top score91.8%Qwen3.6 Max
Models tested62
Last updated2026.09.10
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Alibaba
    Qwen3.6 Max
    2026.04.27 · openrouter-reasoning
    91.8%
  2. 2
    Anthropic
    Claude Fable 5
    2026.06.09 · adaptive-thinking-max with fallback to Opus 4.8
    90.7%
  3. 3
    Alibaba
    Qwen3.5 397B A17B
    2026.02.16 · reasoning-enabled
    90.7%
  4. 4
    Alibaba
    Qwen3.6 35B A3B
    2026.04.27 · reasoning-enabled
    90.2%
  5. 5
    Moonshot AI
    Kimi K3
    2026.07.17 · reasoning-enabled
    90.1%
  6. 6
    OpenAI
    GPT-5.4
    2026.03.05 · high-effort
    89.9%
  7. 7
    Anthropic
    Claude Opus 4.6
    2026.02.04 · extended-thinking
    89.9%
  8. 8
    Anthropic
    Claude Opus 4.1
    2025.08.05 · extended-thinking
    89.8%
  9. 9
    Upstage
    Solar Pro 4
    2026.07.22 · reasoning-enabled
    89.6%
  10. 10
    Anthropic
    Claude Opus 4.7
    2026.04.16 · adaptive-thinking-xhigh
    89.5%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.