DeepSeek-V2.5

DeepSeek
Arena score
1271
Among open models
#97of 148
Among all on Arena
#285
Votes
24,572

Measured on DeepSeek-V2.5, the served version, not a local build. LMArena, CC BY 4.0, published Nov 11, 2026.

Score by category

CategoryScore
  • Coding1301
  • Software & IT1295
  • French1289
  • English1284
  • Legal & government1282
  • Chinese1281
  • Hard prompts, English1280
  • Applied math1274
  • Science1273
  • Long prompts1273
  • Math1271
  • Hard prompts1270
  • Multi-turn1261
  • Medicine & health1259
  • Non-English1253
  • Russian1253
  • Instructions1249
  • Business & finance1248
  • Spanish1248
  • Writing & literature1247
  • Expert1239
  • Creative writing1237
  • Media & sports1230
  • German1226
  • Korean1210
  • Excluding ties1193
  • Japanese1190

Score by size

Plot against
Best for its size
ModelParamsScore
DeepSeek-V2.5236B1271
DeepSeek-V2.5-1210236B1294

About as strong, smaller

Compare

Built on it

Quantizations
8
Fine-tunes
7
Adapters
2
Merges
0

Ways to run it

Most downloaded

BenchClub runs

Not runnable yet

Runs open once a DeepSeek-V2.5 build is registered.

More from DeepSeek

See more

Other 80–300B models

See more