GDPval-AA v2
reported score
1,687
Reported launch score. Public run files and an independent reproduction are still pending.
Kimi K3 benchmarks & scores
67.3
Tentative VCI · #1
Current benchmark data gives Kimi K3 a tentative Vibe Coding Index of 67.3, placing it at rank #1.
Tentative estimate. The 67.3 VCI uses the best current source evidence. Intelligence 58.8, Coding 83.9, Agentic 50.7. Every score and rank will update as fresh results arrive.
Tentative quality
#1/90
67.3
Tentative VCI
Price
#78/91
$6
$/M blended
Current benchmark evidence
Kimi K3 benchmark results
Current available data places Kimi K3 at rank #1.
AA-Briefcase
reported score
1,527
Reported launch score. Public run files and an independent reproduction are still pending.
BrowseComp
reported score out of 100
91.2 / 100
Reported single-agent result with a 1M context window. Independent verification is still pending.
Code Arena WebDev
preliminary Elo
1,679 ±17
Live human-preference WebDev result from 1,757 votes. Arena currently labels this entry Preliminary.
New coding evidence
Reported coding benchmark snapshot
Six reported coding scores, all run with thinking effort marked “max,” supporting the current #1 placement.
DeepSWE
#3 reported
Terminal Bench 2.1
#2 reported
FrontierSWE
#2 reported
Program Bench
#1 reported
Kimi Code Bench 2.0
Internal#2 reported
SWE Marathon
#1 reported
Provenance note, the reported chart is tied to the linked WeChat article and has a public community copy. Public run files and independent reproduction are still pending, Kimi Code Bench 2.0 is explicitly labeled internal.
All Kimi K3 scores and ranks are tentative estimates based on the best currently available benchmark data.
Kimi guide ↗
OpenRouter listing ↗
Suite radar
Suite scores
Intelligence
General reasoning & knowledge (Intelligence Index)
58.8
Coding
Code generation & software tasks (Coding Index)
83.9
Agentic
Multi-step tool use & agentic workflows (Agentic Index)
50.7
TheVibeFather rankings
TentativeKimi K3 practical agentic coding scorecard
This is our tentative algorithmic assessment of where Kimi K3 ranks for common vibe coding and agentic engineering workflows, calculated consistently from the estimated Intelligence, Coding, and Agentic signals above.
Tentative VibeFather rating
6.7
/10
#1 of 90 models
Practical area
Our score
Field rank
Small, well-defined code changes
Precision on scoped edits, fixes, and implementation tasks.
Exceptional
7.6/10
Field rank
#1 / 90
UI and CSS iteration
Front-end implementation with iterative tool-driven refinement.
Exceptional
7.7/10
Field rank
#1 / 90
Routine debugging
Diagnosing failures and turning reasoning into correct code changes.
Strong
7.0/10
Field rank
#1 / 90
Repository-wide refactors
Coordinating larger edits across files while preserving intent.
Strong
6.8/10
Field rank
#1 / 90
Architecture decisions
Reasoning through tradeoffs, constraints, and system-level choices.
Capable
5.6/10
Field rank
#2 / 90
Tool use and workflow execution
Planning and completing multi-step work with tools and feedback loops.
Capable
5.9/10
Field rank
#3 / 90
Long autonomous coding tasks
Sustaining coherent progress across longer agentic engineering runs.
Capable
5.2/10
Field rank
#3 / 90
Overall VibeFather rating
Our complete Vibe Coding Index, expressed on the same 0–10 practical scale.
Strong
6.7/10
Field rank
#1 / 90
1 · Tentative signals
We start with Intelligence, Coding, Agentic benchmark evidence, currently estimated from launch-period results, and preserve missing values as unverified.
2 · Practical lenses
Our code-controlled algorithm blends the native 0–100 signals by workflow and expresses the result on a stable 0–10 scale. Every required input must be present.
3 · Field ranking
Each workflow score is ranked against every model with comparable evidence, so positions update when the benchmark field changes.
The Vibe Coding Index is TheVibeFather’s code-controlled 0–100 composite for practical coding and agentic capability. Every score and rank in this Kimi K3 scorecard is tentative until final source indices arrive. Price and adoption are reported independently and never alter quality. Exact weighting remains proprietary, inputs, missing-data behavior, and per-category ranks are disclosed here. Read the benchmark methodology →
Where Kimi K3 sits
Its logo stays full-color and highlighted while the rest of the field recedes for quick comparison.
Alibaba
Amazon
Anthropic
Cohere
DeepSeek
Google
Inclusionai
Kwaipilot
Meta
Meta Llama
MiniMax
Mistral
Moonshot
NVIDIA
Nex Agi
OpenAI
Stepfun
Tencent
Upstage
Xiaomi
Zhipu
xAI
Hover, tap or focus a model for axis guides · resting dashed lines = medians
20 model(s) hidden — data not yet verified
Company pulse
Latest from Kimi
Up to ten recent public posts from the company, linked to the originals on X.
Loading recent posts…
Public posts are cached briefly for speed.
Model benchmark FAQ
Kimi K3 for vibe coding and agentic engineering
Answers update from this model’s current sourced scores, practical rankings, and verified pricing.
What are the current Kimi K3 benchmark results?
The current tentative profile gives Kimi K3 a 67.3 Vibe Coding Index from Intelligence 58.8, Coding 83.9, and Agentic 50.7. Reported launch results also include 1,687 on GDPval-AA v2 at #3, 1,527 on AA-Briefcase at #2, and 91.2 / 100 on BrowseComp at #1.
What are the reported Kimi K3 coding benchmark scores?
The launch-reported coding chart lists 67.5 / 100 on DeepSWE, 68.3 / 100 on Terminal Bench 2.1, 81.2 / 100 on FrontierSWE, 77.8 / 100 on Program Bench, 72.9 / 100 on the internal Kimi Code Bench 2.0, and 42.0 / 100 on SWE Marathon. Public run files and independent reproduction are still pending.
Where would Kimi K3 rank?
The current multi-source tentative VCI places Kimi K3 at rank #1. Arena Code WebDev currently places it #1 at 1,679 ±17, Arena labels that result preliminary while more votes arrive.
Are the Kimi K3 benchmark scores verified?
The 67.3 VCI, its three component scores, and the #1 placement remain tentative. The calculation uses the best current source evidence and will update automatically as fresh benchmark snapshots arrive.
How much does Kimi K3 cost?
OpenRouter currently lists Kimi K3 at $3 per million input tokens and $15 per million output tokens.
What is confirmed about Kimi K3?
Moonshot confirms a 2.8-trillion-parameter model with native visual understanding and a 1M-token context window. OpenRouter also lists the live model and current pricing.