Kimi K3
currentKimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight multimodal reasoning model for hard coding, knowledge work, and long-running agents. It offers a one-million-token context and strong launch results, but reasoning is always on, output is relatively expensive, and initial API capacity has been limited.
Context window
1.05M tokens
Maximum output
66K tokens
Input price
$3 / 1M
Output price
$15 / 1M
Best for
- Hard coding and large-repository work
- Long-horizon tool-using agents
- Multimodal analysis with a very large context
- Teams interested in open weights at frontier scale
Not ideal for
- Low-cost output-heavy applications
- Tasks that need reasoning disabled
- Latency- or availability-sensitive production systems during capacity constraints
- Fine control over temperature or reasoning effort
PromptHQ 100-word challenge
We give every model the same miniature science-fiction mystery. The response is short enough to read in one chat bubble, while the premise tests creative writing, compression, foreshadowing, causal coherence, and instruction following.
Kimi K3100 wordsWrite a science-fiction mystery in exactly 100 words.
A three-person research station on Europa receives a message from itself, timestamped 11 minutes in the future:
"Do not open the greenhouse door."
The door is opened anyway.
Plant two subtle clues. Give the warning a logical explanation without aliens, time travel, hallucinations, or simulations. The final sentence must change what the reader thought the warning meant.
Output only the story. Do not explain it.
At 04:12, Europa Station’s console woke Bea, Tomás, Ines with a note from itself, stamped 04:23: “Do not open the greenhouse door.” No receiver logged a ping; the print ribbon was warm. Ines saw rime inside the hinge while the greenhouse read summer. Tomás opened it. Air did not rush for one breath. Rows of basil stood black, frozen upright under a dome creaking inward. At 04:23 the clock jumped, erasing eleven minutes of drift after the relay pass. The note was a latch command, misfiled as mail. It had forbidden the door, not the crew, from answering the ice.
Run the same prompt yourself
Open Kimi K3 in Multichat with the evaluation prompt prefilled.
What the response revealed
The response contains exactly 100 whitespace-delimited words.
It outputs only the story and avoids aliens, time travel, hallucinations, and simulations.
Clock drift and a misfiled latch command plausibly explain the future timestamp and apparent message, though the precise function of the command and the final phrase remain ambiguous.
The missing receiver ping, warm print ribbon, internal rime, delayed airflow, and inward-creaking dome all prepare the mechanical and environmental reveal.
The last sentence recasts the warning as an instruction addressed to the door rather than a warning addressed to the crew.
The prose is vivid and compressed, with memorable images and an unusually inventive reinterpretation of the warning.
The prompt was submitted unchanged through Kimi's first-party chat interface. The API route, reasoning setting, token usage, and latency were not recorded. We preserve the response exactly as supplied apart from display rendering.
Performance
Kimi K3 benchmarks
Benchmark scores are sensitive to reasoning effort, harness, tools, token budget, prompt format, sampling, and evaluation date. Scores here retain their source and should not be treated as directly interchangeable unless the underlying setup matches.
DeepSWE v1.1
67.3%
long-horizon coding · Tom's Hardware3
BrowseComp, no context management
90.4%
agentic web research · Tom's Hardware3
Frontend Code Arena
1,679 Elo
front-end generation · Tom's Hardware3
Family position
Kimi K3 positioning
K3 is Moonshot's flagship reasoning and agent model. It prioritizes maximum capability and long-context work; less expensive Kimi models remain better suited to routine workloads.
Kimi K3
This modelFlagship reasoning and agents
$3 input
$15 output
DeepSeek V4 Pro
Lower-cost open rival
$0.43 input
$0.87 output
GLM 5.2
Open-weight coding rival
$0.93 input
$3 output
API pricing
Per million text tokens
Input
$3
Cached input
$0.30
Cache write
$3
Output
$15
Where Kimi K3 stands out
Built for demanding work
Hard coding and large-repository work are central to the model's positioning, rather than an incidental capability.
Long-context capacity
The published context window is 1,048,576 tokens, making the model a candidate for large documents, repositories, and sustained agent state.
Reasoning and tools
Reasoning is supported with max provider setting(s), and the model can participate in tool-using workflows through its available API surface.
Limitations to know
Benchmarks are configuration-sensitive
Scores can move substantially with the harness, tool access, effort setting, token budget, and evaluator. Treat the table as evidence, not a universal ranking.
Context size is not guaranteed recall
A large advertised window does not mean every detail is retrieved reliably at maximum length. Validate representative long-context workloads before deployment.
Product access differs from model capability
PromptHQ and gateway limits may expose fewer modalities, tools, or tokens than the provider's first-party API.
Capabilities and specifications
Knowledge cutoff
Not publicly disclosed
Inputs
text, image, PDF
Reasoning
Supported
Default effort
provider-default
Supported API features
Frequently Asked Questions
Sources
- 1Kimi K3
OpenRouter · gateway model page
- 2Kimi K3 Tech Blog: Open Frontier Intelligence
Moonshot AI · official launch post
- 3Kimi K3 launch benchmark reporting
Tom's Hardware · independent launch coverage
- 4PromptHQ model registry
PromptHQ · internal product configuration