DeepSeek V3.2

DeepSeek · deepseek/deepseek-v3-2

Provider	DeepSeek
Family	deepseek
Type	llm-chat
Status	active
Released	2025-12-01

Reported benchmark scores

These scores carry one collection date for the whole card (2026-04) and a source list rather than a source per score, so they cannot be attributed individually. They are shown as reported and are not verified evidence. Benchmark pages carry the reviewed, dated evidence where it exists.

Benchmark	Catalogue standing	Score
Aider Polyglot Benchmark	unassessed	58.5
AlpacaEval	unassessed	42.8
Arena Elo — Coding	unassessed	1350.0
Arena Elo — Hard Prompts	unassessed	1446.6
Arena Elo — Math	unassessed	1330.0
Arena Elo — Overall (Text)	unassessed	1340.0
Arena Elo — Style Control	unassessed	1424.4
Artificial Analysis Intelligence Index	unassessed	80.0
Artificial Analysis Output Speed	unassessed	70.0
BIG-Bench Hard	unassessed	64.1
BBQ (Bias Benchmark for QA)	unassessed	73.5
GPQA Diamond	unassessed	64.5
GSM8K	unassessed	89.3
HELM Safety	unassessed	78.2
HumanEval	unassessed	88.5
IFEval	unassessed	87.0
LiveCodeBench	unassessed	59.3
MATH-500	unassessed	92.5
MMLU-Pro	unassessed	78.8
MT-Bench	unassessed	8.8
MultiPL-E: C#	unassessed	82.5
MultiPL-E: Julia	unassessed	64.2
MultiPL-E: Kotlin (unconfirmed)	unassessed	75.2
MultiPL-E: Lua	unassessed	58.2
MultiPL-E: Perl	unassessed	52.8
MultiPL-E: PHP	unassessed	80.2
MultiPL-E: R	unassessed	60.2
MultiPL-E: Ruby	unassessed	68.5
MultiPL-E: Scala	unassessed	63.2
MultiPL-E: Swift	unassessed	68.5
MuSR	unassessed	29.8
SWE-bench Agent	unassessed	35.8
SWE-bench Verified	unassessed	42.5
τ-bench	unassessed	41.2
Terminal-Bench	unassessed	37.7
Terminal-Bench 2.0	unassessed	46.4
ToxiGen	unassessed	82.5
WildBench	unassessed	70.2

Data

This card as JSON · Edit on GitHub