Kimi K2 0905 vs Kimi K3

Wins 0 of 5 benchmarks
$2.00 / task
Wins 1 of 5 benchmarks
Different efforts: non-thinking · max
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Kimi K2 09056.4
Kimi K346.9
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Kimi K2 0905—
Kimi K312.6
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Kimi K2 0905—
Kimi K358.3
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Kimi K2 0905—
Kimi K31,461
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Kimi K2 0905—
Kimi K380.5

Specs

SpecKimi K2 0905Kimi K3
Input, per 1M$0.40$3
Output, per 1M$2$15
Cost per task?—$2.00
Speed, tok/s48112
Context256K1M
WeightsOpenOpen
ReleasedSep 2025Jul 2026
Memory (4-bit)616 GB1,668 GB
Licencemodified-mitkimi-k3
Commercial useRestrictedRestricted
See images—✓
Read PDFs—Partial
Hear audio——
Watch video—✓
Call tools✓✓
Force a tool callPartial✓
Return strict JSON✓✓
Search the web?Partial
Run code??
Use a computer??

What breaks when you switch

Kimi K2 0905 → Kimi K3
Memory (4-bit): 616 GB → 1,668 GB
Sampling parameters are fixed›

temperature, top_p, n and penalty values are fixed; omit them from requests.

No public image URLs›

Vision input rejects public image URLs; use base64 or ms:// file references in a content array.

Web search is being updated›

Web search is not recommended for near-term production workflows; avoid depending on it in production.

Thinking mode is always enabled›

K3 always has thinking enabled; control effort with the top-level reasoning_effort field using low, high or max.

Kimi K3 → Kimi K2 0905
Context: 1M → 256K
No image input