Moonshot AI · China
Kimi K3
kimi-k3
NovedadOpen-weightMás barato
Familia Kimi. Contexto largo y enjambres de agentes.
Input / 1M
$2.28~
aproximado
Output / 1M
$11.4~
aproximado
Contexto
1.048576M
tokens
SWE-bench
—
Pro / Verified
Fortalezas
- Ventana de 1M+ tokens: útil en repos grandes y multiarchivo.
- Open-weight: puedes hospedarlo y auditar los pesos.
- Pensado para flujos agénticos de coding.
- Multimodal: text, image.
Matices
- Sin SWE-bench publicado en este catálogo.
- Precio aproximado: confirma siempre en la web oficial.
Casos de uso
- Agentes de código y tareas de ingeniería de software.
- Repositorios grandes y trabajo multiarchivo.
- Despliegue local o en infraestructura propia.
Límites y ficha
- Contexto
- 1.048576M
- Modalidades
- text, image
- Licencia
- Open-weight
- Parámetros
- 2800B
- Lanzamiento
- jul 2026
- Origen
- China
Benchmarks
- SWE-bench
- —
- LiveCodeBench
- —
- HumanEval
- —
Notas
MoE de 2.8T parámetros, open-weight. TerminalBench 85.0%, SciCode 58.7%, líder en Design Arena (código Elo >1400). Contexto 1M y razonamiento profundo.
Ejemplo de uso
SDK Python (OpenAI-compatible)
from openai import OpenAI
client = OpenAI(
base_url="https://api.moonshot.cn/v1",
api_key="$MOONSHOT_API_KEY",
)
resp = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Escribe un script para scrapear esta web..."}],
)
print(resp.choices[0].message.content)Historial de la ficha
- 2026-09-29Ficha refrescada con precios, contexto y benchmarks del Intelligence Index.
Actualizado en catálogo el 29 sept 2026. Confirma precio y límites en la web oficial.