Glimmer Index
Plus $4/mo

DeepSeek-V3

DeepSeek · Released Dec 26, 2024 · Open weights

Open DeepSeek MoE chat model for coding, math, and general reasoning

Strongest at
Coding
Better than 34% of the past year's models · one benchmark so far
Context window
131.1K
tokens in · 8.2K out
Since the last one
112 days
after DeepSeek-V2.5
Lab announcement ↗Open in the timeline

Capability profile

Scored against 91 models released in the past year. 50 is the average model; further out is better.

This modelMedian of the past year
5090Median of the past yearCoding: 52 · Better than 34%Reasoning: 25 · Better than 4%Math: 34 · Better than 4%Science: 45 · Better than 2%CodingReasoningMathScience
Coding52Better than 34% · 1 test
Reasoning25Better than 4% · 1 test
Math34Better than 4% · 1 test
Science45Better than 2% · 1 test

No results yet: Writing, Reading & documents, Visual reasoning, Computer use

Where these scores come from

Each capability combines the published benchmarks listed here, rescaled so 50 is the reference average and 15 points is one standard deviation. A hollow dot rests on a single benchmark.

Since DeepSeek-V2.5

8 releases in the DeepSeek V4 Preview line.

DeepSeek-V2.5Sep 5, 2024112dDeepSeek-V3Dec 26, 202488dDeepSeek V3 0324Mar 24, 2025150dDeepSeek V3.1Aug 21, 20250dDeepSeek-V3.1Aug 21, 202539dDeepSeek-V3.2-ExpSep 29, 202563dDeepSeek V3.2Dec 1, 2025
Released112 days later

A newer release is out: DeepSeek V4 Preview.

ECI STANDING#154 of 270
SCORE132.3128.06–135.23 CI
FRONTIER LAGLaunched 3.4 months behindScores about what o1-mini did in September 2024.
SCORES LIKE THESE MODELS
Magistral Small 1.0Mistral AI · Jun 2025133.07
Qwen2.5-MaxAlibaba · Jan 2025132.48
Llama 4 MaverickMeta AI · Apr 2025132.22
Gemini 1.5 Pro (Sept 2024)Google DeepMind · Sep 2024131.72
Full Epoch Capabilities Index ranking ↗

Epoch benchmark boards

Best result on each board it has run. The tick marks the leader.

MMLU
87%#3 of 127
ARC AI2
95%#2 of 76
BBH
88%#2 of 45
HellaSwag
89%#4 of 75
Winogrande
85%#6 of 77
PIQA
85%#6 of 59
TriviaQA
83%#10 of 38
MATH level 5
65%#42 of 98
2 more boards
METR Time Horizons
47%#28 of 42
SimpleBench
19%#85 of 89

DeepSeek's past year

9 release days in the past year, about one every 41 days. The next is expected around Nov 28, 2026. Forecasts ↗

NovDecJanFebMarAprMayJunJulAugSepOctNov

Specifications

RELEASE DATEDec 26, 2024
WEIGHTSOpen weights
CONTEXT WINDOW131.1K
MAX OUTPUT8.2K
REASONINGNot listed
TOOL CALLINGNot listed
INPUTStext
OUTPUTStext

Models.dev record ↗

Specifications and prices from Models.dev; availability and limits may differ by provider. Benchmarks from Epoch AI (CC BY 4.0), LiveBench and the other sources named above.