name = "Kimi K3"
description = "Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work"
family = "kimi-k3"
release_date = "2026-07-16"
last_updated = "2026-07-16"
attachment = true
reasoning = true
temperature = false
tool_call = true
structured_output = true
open_weights = true

[limit]
context = 1_048_576
output = 131_072

[modalities]
input = ["text", "image", "video"]
output = ["text"]

[[benchmarks]]
name = "DeepSWE"
score = 67.5
metric = "resolve rate"
variant = "max effort"
harness = "Kimi Code"
version = "1.1"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "Terminal-Bench"
score = 88.3
metric = "accuracy"
variant = "max effort"
harness = "Kimi Code"
version = "2.1"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "FrontierSWE"
score = 81.2
metric = "dominance score"
variant = "max effort"
harness = "Kimi Code"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "Program Bench"
score = 77.8
metric = "score"
variant = "max effort"
harness = "Kimi Code"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "SWE Marathon"
score = 42.0
metric = "resolve rate"
variant = "max effort"
harness = "Claude Code"
version = "1.1"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "GDPval-AA"
score = 1668
metric = "Elo"
variant = "max effort"
version = "v2"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "AA-Briefcase"
score = 1548
metric = "Elo"
variant = "max effort"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "AutomationBench"
score = 30.8
metric = "success rate"
variant = "max effort"
dataset = "600-task public subset"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "JobBench"
score = 52.9
metric = "score"
variant = "max effort"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "SpreadsheetBench"
score = 34.8
metric = "score"
variant = "max effort"
harness = "Claude Code"
version = "2"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "BrowseComp"
score = 91.2
metric = "accuracy"
variant = "max effort, context compaction"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 91.3
metric = "accuracy"
variant = "max effort, with tools"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"

[[benchmarks]]
name = "ZeroBench"
score = 41.0
metric = "pass@5"
variant = "max effort, with tools"
source = "https://www.kimi.com/blog/kimi-k3"
date = "2026-07-16"