Kimi K2.5

Training compute

5.8×10²⁴ FLOP

Parameters

1T

Published

Feb 2, 2026

Kimi K2.5 is an AI model developed by Moonshot (China), first published in February 2026. It works in the language domain, on tasks such as language modeling/generation.

Training it took an estimated 5.8×10²⁴ FLOP of compute. The model has 1,040,000,000,000 parameters. Training ran on NVIDIA H800 SXM5.

Access: Open weights (unrestricted). Its weights are openly available. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Moonshot
Country of organization: China
Domain: Language
Task: Language modeling/generation
Training compute: 5.8×10²⁴ FLOP
Parameters: 1,040,000,000,000
Training hardware: NVIDIA H800 SXM5
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Likely

Benchmark results

01	WebDev Arena	Arena Score	1430.6
02	Vending-Bench 2	Score	1198.5
03	ALE Bench	Performance	821.6
04	Epoch Capabilities Index	ECI Score	148.3
05	OSWorld	Score	63.3
06	Fictionlivebench	120k token score	0.78
07	SWE-bench Verified	mean_score	0.74
08	ARC-AGI	Score	0.65
09	SimpleBench	Score (AVG@5)	0.47
10	WeirdML	Accuracy	0.46
11	Terminal-Bench	Accuracy mean	0.43
12	Cursorbench	Score	0.32
13	Frontierswe	Dominance	0.27
14	Humanity's Last Exam	Accuracy	0.24
15	CL Bench	Overall	0.19
16	Apex Agents	Pass@1 score	0.14
17	CL Bench Life	Overall	0.13
18	ARC-AGI-2	Score	0.12
19	Posttrainbench	Average (%)	0.1
20	Enigma Eval	Accuracy	0.03