Atlas

Models

← All models

DeepSeek-V2.5

DeepSeek · 2024-09-05 · 235.7B parameters

DeepSeek-V2.5 is the open-weight merger of DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724, released September 5, 2024. It combines general conversational and coding capabilities with improved human-preference alignment, writing, instruction following, and code performance; it retains the DeepSeek-V2 MoE architecture, marketed as 236B total / 21B activated, and a documented 128K-token context.

Benchmark scores

BenchmarkScore
Aider Polyglot17.8
Arena Text — Business, Management & Financial Ops (No Style Control)1248
Arena Text — Chinese (No Style Control)1280
Arena Text — Creative Writing (No Style Control)1237
Arena Text — English (No Style Control)1284
Arena Text — Entertainment, Sports & Media (No Style Control)1231
Arena Text — Exclude Ties (No Style Control)1193
Arena Text — Exclude Ties (Style Controlled)1244
Arena Text — Expert (No Style Control)1239
Arena Text — French (No Style Control)1288
Arena Text — French (Style Controlled)1325
Arena Text — German (No Style Control)1226
Arena Text — Hard Prompts (No Style Control)1271
Arena Text — Hard Prompts English (No Style Control)1282
Arena Text — Instruction Following (No Style Control)1250
Arena Text — Japanese (No Style Control)1190
Arena Text — Korean (No Style Control)1210
Arena Text — Legal & Government (No Style Control)1282
Arena Text — Life, Physical & Social Science (No Style Control)1274
Arena Text — Longer Query (No Style Control)1273
Arena Text — Medicine & Healthcare (No Style Control)1258
Arena Text — Multi-turn (No Style Control)1261
Arena Text — Non-English (No Style Control)1253
Arena Text — Non-English (Style Controlled)1290
Arena Text — Occupational Mathematical (No Style Control)1274
Arena Text — Russian (No Style Control)1252
Arena Text — Software & IT Services (No Style Control)1296
Arena Text — Spanish (No Style Control)1247
Arena Text — Writing, Literature & Language (No Style Control)1247
Artificial Analysis Intelligence Index6.6
Loading Atlas data…