Skip to content
fedi.software

DeepSeek V4 Pro 0423 in fedi-index

The score in every slice, what it is made of, value, local run and history.

fedi-index · General

A 1,427 ±5

Percentile 82 · #54

People prefer / Solves tasks

1,437 / 1,415

Human votes vs benchmarks, Elo-eq

Value

+33

$1.91 per 1M tokens, median of 49 providers

On your hardware

Does not fit the reference setups.

Score decomposition

A General 1,427 [1,422–1,432] Percentile 82 · #54 · components: 9
Source Component Value Elo-eq Weight Measured Variant
Epoch AI eci 149 ±0.87 1,431 ±7 34.5% 2026-04-24 —
LMArena text/overall 1,451 ±1.98 1,451 ±2 15.6% 2026-10-02 —
LMArena text/hard_prompts 1,461 ±2.33 1,450 ±2 15.6% 2026-10-02 —
LMArena text/expert 1,463 ±4.29 1,446 ±4 7.5% 2026-10-02 —
LMArena text/instruction_following 1,447 ±5.25 1,458 ±5 7.2% 2026-10-02 high
Epoch AI simpleqa_verified decay ×0.61 0.47 ±0.02 1,445 ±5 3.8% 2026-08-27 max
LMArena text/multi_turn 1,467 ±3.57 1,460 ±3 3.8% 2026-10-02 —
LMArena text/longer_query 1,458 ±4.71 1,454 ±4 3.7% 2026-10-02 high
Epoch AI gpqa_diamond decay ×0.26 0.91 ±0.02 1,463 ±10 1.3% 2026-08-06 high

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Coding 1,439 [1,434–1,444] Percentile 90 · #30 · components: 3
Source Component Value Elo-eq Weight Measured Variant
LMArena webdev/overall 1,583 ±4.76 1,473 ±1 40.6% 2026-10-01 high
LMArena text/coding 1,469 ±5.79 1,443 ±5 36.6% 2026-10-02 high
Epoch AI swe_bench_verified decay ×0.37 0.78 ±0.02 1,472 ±10 14.2% 2026-06-18 max

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Agents 1,435 [1,428–1,442] Percentile 54 · #27 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena agent/overall 0.01 ±0.01 1,461 ±4 90.0% 2026-10-02 high-0813

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Math 1,413 [1,407–1,420] Percentile 79 · #54 · components: 3
Source Component Value Elo-eq Weight Measured Variant
LMArena text/math 1,455 ±6.01 1,446 ±6 49.7% 2026-10-02 high
Epoch AI frontiermath_t1_3 decay ×0.86 0.45 ±0.03 1,419 ±6 27.8% 2026-06-17 max
Epoch AI frontiermath_t4 decay ×0.86 0.02 ±0.02 1,396 ±4 15.4% 2026-06-17 max

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Writing 1,452 [1,446–1,457] Percentile 87 · #36 · components: 2 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/creative_writing 1,446 ±3.69 1,475 ±4 62.9% 2026-10-02 —
LMArena text/longer_query 1,458 ±4.71 1,454 ±4 30.8% 2026-10-02 high

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Language: 繁體中文 1,427 [1,419–1,435] Percentile 76 · #54 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/chinese 1,486 ±5.57 1,442 ±4 93.6% 2026-10-02 high

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Language: Deutsch 1,432 [1,417–1,447] Percentile 80 · #30 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/german 1,457 ±9.86 1,451 ±8 92.4% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Language: English 1,436 [1,430–1,440] Percentile 86 · #39 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/english 1,459 ±2.60 1,450 ±3 93.9% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Language: Español 1,434 [1,421–1,447] Percentile 80 · #31 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/spanish 1,458 ±7.78 1,452 ±7 92.9% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

B Language: Français 1,435 [1,422–1,447] Percentile 78 · #31 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/french 1,472 ±7.44 1,452 ±7 93.0% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Language: 日本語 1,441 [1,425–1,458] Percentile 84 · #21 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/japanese 1,447 ±12 1,462 ±9 92.2% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Language: Polski 1,441 [1,427–1,455] Percentile 84 · #22 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/polish 1,462 ±9.13 1,459 ±8 92.7% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

A Language: Русский 1,439 [1,432–1,446] Percentile 84 · #39 · components: 1 provisional
Source Component Value Elo-eq Weight Measured Variant
LMArena text/russian 1,453 ±4.18 1,455 ±4 93.8% 2026-10-02 —

Weight is the share of the component in the score (the rest is the prior that pulls models with little data towards the population median). Methodology

History

fedi-index by month
The same in a table
Model 2026-10
DeepSeek V4 Pro 0423 1,427 A

Badge

Show the model's fedi-index on your site or in a README. The badge updates itself; it links to this page.

fedi-index: DeepSeek V4 Pro 0423

Data CC BY 4.0: keep the link to fedi.software.