Agenci AI · karta decyzji
Kimi K3
Moonshot AI
Kimi K3 is an open source 2.8 trillion parameter model with 1M context window, native vision, and continuous reasoning, designed for coding, agentic tasks, and knowledge work.
Dostęp
Freemium
Cena od
0
Trial
unknown
Sprawdzone
9.08.2026
W skrócie
Czy to narzędzie pasuje do Twojego zadania?
Free tier on kimi.com, subscriptions $19-$199/month, API $3 input/$15 output per million tokens, self-hosting with free weights.
Zastosowania
- coding
- long-duration agentic sessions
- knowledge work
- code repository analysis
- bug fixing
Najważniejsze funkcjonalności
- Open source model with 2.8 trillion parameters
- 1 million token context window
- Native vision (image reading)
- Continuous reasoning with adjustable effort
- Tool use and terminal control
- Code repository scanning and bug fixing from logs/screenshots
- Quantization-aware training (MXFP4 weights, MXFP8 activations)
- OpenAI/Anthropic-compatible API
- Self-hosting with free weights
Darmowy plan i trial
Darmowy dostęp: Kimi K3 is a free-to-use open source model on kimi.com, with paid subscriptions for intensive agentic sessions and API pricing per token.
Trial: unknown
Cennik: Free tier on kimi.com, subscriptions $19-$199/month, API $3 input/$15 output per million tokens, self-hosting with free weights.
Plany i limity
| Plan | Cena | Dostęp | Limit / zawartość | Trial / karta |
|---|---|---|---|---|
| Free | 0 USD | free | unknown | unknown · karta: unknown |
| Subscription | 19-199 USD / month | subscription | unknown | unknown · karta: unknown |
| API | 3 input, 15 output per million tokens USD / per_token | api | unknown | unknown · karta: unknown |
| Self-hosted | free weights, hardware at user cost unknown | self-hosted | unknown | unknown · karta: unknown |
Dane i bezpieczeństwo
- Prywatność
- unknown
- Trening na danych
- unknown
- Region danych
- unknown
- RODO / GDPR
- unknown
- Bezpieczeństwo
- unknown
Materiały i dowody
Kimi K3 achieves competitive scores on benchmarks like PostTrainBench, Kimi Code Bench 2.0, Deep Swe, Apex Agents, GPQA, and HLE, often matching or exceeding proprietary models. The tech report provides detailed evaluation.
Badania, benchmarki i raporty
Tytuł materiału producenta nie jest niezależnym dowodem skuteczności. Źródła oznaczamy według ich pochodzenia.