llmrun
Coding

GBA Eval Leaderboard

GBA Eval, from Mechanize, gives a coding agent a long task: build a Game Boy Advance emulator from scratch in Rust and WebAssembly. The emulator is graded against a reference emulator on replayed gameplay, hardware test ROMs and audio; the 0–1 overall score is shown as redistributed by Epoch AI.

Source: epoch7 open models ranked+16 proprietaryData through Jul 2026

Open modelsAll models

All models ranked on GBA Eval

Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.

#ModelScoreVRAM @ Q4
1Claude Opus 5 (unspecified) · proprietary
79.6%
API only
2Claude Fable 5 (unspecified) · proprietary
74.5%
API only
3Claude Opus 4.8 (unspecified) · proprietary
70.9%
API only
4Grok 4.5 (unspecified) · proprietary
65.4%
API only
5Claude Sonnet 5 (unspecified) · proprietary
65.3%
API only
6GPT 5.5 (unspecified) · proprietary
53.2%
API only
7GPT 5.6 Sol (unspecified) · proprietary
52.6%
API only
8Claude Sonnet 4.6 (unspecified) · proprietary
48.8%
API only
9Kimi K3 · 2779.9B
48.4%
~1674 GB
10GPT 5.4 (Mar 05, 2026, unspecified) · proprietary
45.1%
API only
11Claude Opus 4.6 (unspecified) · proprietary
44.1%
API only
12Claude Opus 4.7 (unspecified) · proprietary
43.8%
API only
13Muse Spark 1.1 · proprietary
7.9%
API only
14Gemini 3.5 Flash (unspecified) · proprietary
6.7%
API only
15Grok Build 0.1 · proprietary
2.4%
API only
16MiniMax M3 · 427.0B
0.9%
~257 GB
17Kimi K2.6 · 1026.9B
0.9%
~620 GB
18Gemini 3.1 Pro Preview · proprietary
0.9%
API only
19Kimi K2.7 Code · 1026.9B
0.9%
~620 GB
20Qwen3.7 Max · proprietary
0.4%
API only
21GLM 5.2 · 753.3B
0.0%
~456 GB
22GLM 5.1 · 753.9B
0.0%
~457 GB
23MiniMax M2.7 · 228.7B
0.0%
~138 GB

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

229B2.8Tmodel size (log scale) →48.4%0.0%Kimi K2.6 · 1T · 0.9%Kimi K2.7 Code · 1T · 0.9%GLM 5.2 · 753B · 0.0%GLM 5.1 · 754B · 0.0%MiniMax M2.7 · 229B · 0.0%MiniMax M2.7MiniMax M3 · 427B · 0.9%MiniMax M3Kimi K3 · 2.8T · 48.4%Kimi K3
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • MiniMax M2.7, 229B, score 0.0% — on the efficiency frontier (best score at its size or smaller).
  • MiniMax M3, 427B, score 0.9% — on the efficiency frontier (best score at its size or smaller).
  • Kimi K3, 2.8T, score 48.4% — on the efficiency frontier (best score at its size or smaller).

GBA Eval: frequently asked questions

What is the best open LLM on GBA Eval?
Kimi K3 is the top open model on GBA Eval, scoring 48.4%. Among all models tested — including proprietary ones — it ranks #9. The top model overall is Claude Opus 5 (unspecified) (Anthropic) at 79.6%.
Can open models match proprietary models on GBA Eval?
Not quite on GBA Eval: the strongest proprietary model (Claude Opus 5 (unspecified)) scores 79.6%, ahead of the best open model (Kimi K3) at 48.4% — but you can run the open one yourself.

Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.