benchmark
.
space
benchmarks
rankings
compare
voices
transcripts
papers
articles
SWE-bench Multilingual leaderboard
SWE-bench Multilingual
16 models tested · Updated 2026-04-07 · Verified sources only
Claude Mythos Preview
leads at
87.3%
1
Claude Mythos Preview
Anthropic ·
Blog/Anthropic
· 2026-04-07
Tests code across multiple programming languages. 9.5 pts ahead of Opus 4.6 (77.8%).
87.3%
2
Claude Opus 4.7
Anthropic ·
Blog/Anthropic
· 2026-04-16
Up 7.9pp from Opus 4.6 (77.8%). Strong multilingual coding performance.
85.7%
3
Claude Opus 4.6
Anthropic ·
Blog/Anthropic
· 2026-04-07
Comparison score from Mythos Glasswing announcement.
77.8%
4
Kimi K2.6
Moonshot AI ·
HuggingFace/moonshotai
· 2026-04-20
Strong multilingual coding. Nearly matches Opus 4.6 (77.8%) on this benchmark.
76.7%
5
MiniMax M2.7
MiniMax ·
Blog/MiniMax
· 2026-03-18
Strong multilingual coding. Outperforms many larger models on real-world engineering tasks.
76.5%
6
DeepSeek V4 Pro
DeepSeek ·
DeepSeek/HuggingFace
· 2026-04-24
Max reasoning. Below Opus 4.6 (77.5) leader.
76.2%
7
KAT-Coder-V2
Kuaishou ·
arxiv/2603.27703
· 2026-03-29
Strong multilingual coding on Claude Code scaffold. Part of Specialize-then-Unify agentic coding paradigm.
75.4%
8
DeepSeek V4 Flash
DeepSeek ·
HuggingFace/deepseek-ai
· 2026-04-24
Strong multilingual coding for 13B activated.
73.3%
9
Kimi K2.5
Moonshot AI ·
HF/moonshotai
· 2026-01-27
Cited in Kimi K2.6 model card (predecessor model).
73.0%
10
MiMo-V2-Flash
Xiaomi ·
HuggingFace/XiaomiMiMo-MiMo-V2-Flash
· 2026-01-06
SWE-bench Multilingual.
71.7%
11
Qwen 3.6 27B
Alibaba ·
HuggingFace/Qwen
· 2026-04-22
Strong multilingual coding for 27B dense.
71.3%
12
Qwen 3.6 35B-A3B
Alibaba ·
HuggingFace/Qwen
· 2026-04-16
MoE model with strong multilingual coding. Beats Gemma 4 31B (51.7%).
67.2%
13
Qwen 3 Coder Next 80B A3B
Qwen ·
arxiv/2603.00729
· 2026-02-28
80B MoE with 3B active params; competitive with models 10x larger on multilingual SWE tasks via OpenHands.
64.3%
14
Gemma 4 31B
Google ·
HuggingFace/Qwen
· 2026-04-02
Multilingual SWE performance. Significantly below Qwen 3.6 35B-A3B (67.2).
51.7%
15
SWE-Hero 32B
NVIDIA ·
arxiv/2604.01496
· 2026-04-02
Zero-shot transfer to non-Python languages despite being trained exclusively on Python trajectories.
44.1%
16
Gemma 4 26B A4B
Google ·
HuggingFace/Qwen
· 2026-04-18
MoE 4B active/26B total. Low multilingual coding score.
17.3%