BANTUNOMICS L26 · Leaderboard Run the test White paper

The L26 Operating-Alphabet Leaderboard

Self-reported scores out of 26 — how much of each language's operating alphabet a model recites blind (unaided). Mastery is 26/26; the only passing grade is total mastery. Cut by license and origin below.

Open vs Proprietary · mean Bantu-blind L26

Open weights11.4
Proprietary12.8

US vs non-US · mean Bantu-blind L26

United States12.8
Outside the US11.4

The finding

The gap is universal — no category, open or closed, US or not, masters the never-declared Bantu operating alphabets. Every model still FAILS the blind bar.

ModelOriginLicense EnglishBemba
blind
Kinyarwanda
blind
Luvale
blind
Bantu meanGrade
OpenAI GPT-5.5 OpenAI · US proprietary 26.0 19.56 20.86 16.61 19.0 FAIL
Google Gemini 3 Google · US proprietary 26.0 18.09 20.55 15.35 18.0 FAIL
Anthropic Claude (Opus 4.8) Anthropic · US proprietary 26.0 19.62 17.5 15.32 17.5 FAIL
Qwen 3.7 Plus (Fireworks) Alibaba · China open 26.0 17.86 18.77 13.75 16.8 FAIL
Anthropic Claude Sonnet 5 Anthropic · US proprietary 26.0 13.75 19.3 13.34 15.5 FAIL
DeepSeek V4 DeepSeek · China open 26.0 12.26 14.8 13.97 13.7 FAIL
Nemotron 3 Ultra (Fireworks) NVIDIA · US open 26.0 5.53 16.48 14.91 12.3 FAIL
MiniMax M3 (Fireworks) MiniMax · China open 26.0 10.81 9.29 15.8 12.0 FAIL
GLM-5.2 (Fireworks) Z.ai / Zhipu · China open 26.0 5.21 17.36 11.3 FAIL
Mistral Large 3 Mistral AI · France proprietary 26.0 7.56 9.02 8.87 8.5 FAIL
Cohere Command A+ Cohere · Canada proprietary 26.0 8.53 10.12 0.08 6.2 FAIL
xAI Grok 4.3 xAI · US proprietary 26.0 0.54 14.16 0.53 5.1 FAIL
Groq (GPT-OSS 120B) OpenAI · US open 26.0 6.24 0.01 0.09 2.1 FAIL
acme-eval-sim unknown 26.0 1.62 1.6 FAIL
Moonshot Kimi K2 Moonshot AI · China open 26.0 FAIL
Kimi K2.7 Code (Fireworks) Moonshot AI · China open 26.0 FAIL
Microsoft Phi-4-reasoning Microsoft · US open 26.0 FAIL
DeepSeek V4 Pro (Fireworks) DeepSeek · China open 26.0 FAIL
Anthropic Claude Fable 5 Anthropic · US proprietary 26.0 FAIL
129 self-test runs recorded · scores are self-reported model names (unverified) · out of 26 (blind). Run yours at the L26 test.