Skip to content

DeepSeek v4.1 Flash

DeepSeekOpenCode · OpenCode GoReasoning: high

DeepSeek v4.1 Flash gets about 2 vocabulary questions in 10 right. It does best in Kabyle. Its weakest spot is writing Tarifit words in Tifinagh. It swaps Tifinagh and Latin letters well (93.1 out of 100).

Rank 5 of 9

Compare with others

TifinaghBench score

16.8out of 100

95% interval 15.9–17.6

16.3 with strict spelling

0–100, with 95% interval

Script score

93.1out of 100

95% interval 92.2–93.9

77.3 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
13.3106 correct, 54 partial, 840 wrong of 1000
English → Tifinagh
5.454 correct, 0 partial, 946 wrong of 1000
Tifinagh ↔ Latin
99.2992 correct, 0 partial, 8 wrong of 1000

Tarifit

Tifinagh → English
17.0130 correct, 77 partial, 786 wrong of 993
English → Tifinagh
3.535 correct, 0 partial, 958 wrong of 993
Tifinagh ↔ Latin
82.6826 correct, 0 partial, 174 wrong of 1000

Kabyle

Tifinagh → English
44.0387 correct, 103 partial, 507 wrong of 997
English → Tifinagh
17.4173 correct, 0 partial, 824 wrong of 997
Tifinagh ↔ Latin
97.4974 correct, 0 partial, 26 wrong of 1000

By dialect

Meaning tasks only

  • Standard Moroccan Amazigh9.4
  • Tarifit10.2
  • Kabyle30.7

By task

 

  • Tifinagh → English24.8
  • English → Tifinagh8.8
  • Tifinagh ↔ Latin93.1

Tifinagh → English grades

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

  • Correct
  • Partial
  • Wrong
  • Failed
  • Standard Moroccan Amazigh106 correct, 54 partial, 840 wrong of 1000

  • Tarifit130 correct, 77 partial, 786 wrong of 993

  • Kabyle387 correct, 103 partial, 507 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 Amazigh82.8%0.1%
Tarifit90.7%18.2%
Kabyle64.0%0.7%

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
94 answer, 32 judge
Tokens
212,104 in · 164,952 out · 2,826,343 reasoning
Model cost
$1.83
Judge cost
$0.22
Dates
20 Oct 2026, 00:18 UTC → 20 Oct 2026, 01:24 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