Skip to content

GLM 5.3 Flash

Zhipu AIOpenCode · OpenCode GoReasoning: high

GLM 5.3 Flash gets about 1 vocabulary question in 10 right. It does best in Kabyle. Its weakest spot is writing Tarifit words in Tifinagh. It still makes mistakes swapping Tifinagh and Latin letters (80.6 out of 100).

Rank 6 of 9Tied with Claude Sonnet 5.5

Compare with others

TifinaghBench score

12.0out of 100

95% interval 11.3–12.7

11.7 with strict spelling

0–100, with 95% interval

Script score

80.6out of 100

95% interval 79.3–81.9

65.4 with strict spelling

Transliteration score, 0–100

Scores by dialect and task

Share of the score earned in each cell, 0–100. The transliteration column measures the script, not the vocabulary.

Standard Moroccan Amazigh

Tifinagh → English
9.468 correct, 51 partial, 881 wrong of 1000
English → Tifinagh
2.626 correct, 0 partial, 974 wrong of 1000
Tifinagh ↔ Latin
90.7907 correct, 0 partial, 93 wrong of 1000

Tarifit

Tifinagh → English
12.086 correct, 66 partial, 841 wrong of 993
English → Tifinagh
1.919 correct, 0 partial, 974 wrong of 993
Tifinagh ↔ Latin
60.6606 correct, 0 partial, 394 wrong of 1000

Kabyle

Tifinagh → English
35.2294 correct, 114 partial, 589 wrong of 997
English → Tifinagh
10.9109 correct, 0 partial, 888 wrong of 997
Tifinagh ↔ Latin
90.5905 correct, 0 partial, 95 wrong of 1000

By dialect

Meaning tasks only

  • Standard Moroccan Amazigh6.0
  • Tarifit6.9
  • Kabyle23.1

By task

 

  • Tifinagh → English18.8
  • English → Tifinagh5.1
  • Tifinagh ↔ Latin80.6

Tifinagh → English grades

Free-text meanings, graded by the judge or matched exactly.

  • Correct
  • Partial
  • Wrong
  • Failed
  • Standard Moroccan Amazigh68 correct, 51 partial, 881 wrong of 1000

  • Tarifit86 correct, 66 partial, 841 wrong of 993

  • Kabyle294 correct, 114 partial, 589 wrong of 997

Character error rate

For exact-match answers: the smallest edit distance to an accepted answer, divided by its length. Lower is better; it shows near misses that a 0 score hides.

English → TifinaghTifinagh ↔ Latin
Standard Moroccan Amazigh92.7%3.7%
Tarifit95.4%25.4%
Kabyle74.5%3.1%

The run

Setup
Official setup
Harness
OpenCode 2.0.32@opencode-ai/client 0.0.0-beta-19271
Provider
OpenCode Go
Judge
Claude Haiku 5.5 (Claude Code)
Sample
1,000 words per dialect, seed 1
Batch size
100
Sessions
102 answer, 30 judge
Tokens
212,948 in · 226,739 out · 1,213,801 reasoning
Model cost
$0.75
Judge cost
$0.23
Dates
20 Oct 2026, 12:34 UTC → 20 Oct 2026, 13:51 UTC
CLI version
0.1.0
Dataset checksums

SHA-256 of each dictionary file, so the run can be checked once the datasets are released.

  • Standard Moroccan Amazigh633505d9ccd2244f9a53b64f56bd153c428cc3dbd7229066845404bf3d7c5060
  • Tarifitd144a5619aa31cc06ffcc4eb26c361c65ca2c71451dfdfbea7fa134206481c4f
  • Kabyleed3a9882ac0f4ef10630917a302fd7d9ff6ac706654bb317b94c715f6b4cf9ea

Run folder on GitHub