Kimi K3 API — Moonshot Flagship Quickstart
Kimi K3 is Moonshot AI's July 2026 flagship: ~2.8 trillion parameters, 1M-token context, native vision — an open 3T-class model for long-horizon coding, knowledge work, and multi-agent systems.
Direct answer (July 2026)
Current flagship = Kimi K3 (not K2 / moonshot-v1-only positioning). Use model="kimi-k3" (confirm dashboard alias) against https://api.chinamodelapi.com/v1. API product docs: Moonshot / platform.kimi.ai; open weights rollout was scheduled around late July 2026.
Sources: CNBC / VentureBeat / Moonshot launch coverage (mid-July 2026); X discussion of weight drop ~2026-07-27.
Specs that matter for builders
- ~2.8T parameters — largest open-class Chinese frontier narrative in July 2026 coverage
- 1M-token context — multi-file repos, research packs, long agent trajectories
- Native vision — multimodal input in the K3 generation
- Prior family — K2.7 Code / K2.x still relevant for cheaper coding workloads
When to pick K3 vs others
| Need | Prefer |
|---|---|
| Max long-horizon agent / huge context | kimi-k3 |
| Open-weight coding harness, self-host | glm-5.2 (weights) or K3 when weights live |
| Cheapest short chat / RAG answer | deepseek-v4-pro / glm flash tiers |
| Multilingual general + 1M Qwen stack | qwen-plus / qwen-max family |
Kimi lineup & pricing (August 2026)
Two Kimi generations are worth routing to today. K3 is the frontier pick; K2.5 is the value pick for agent loops that don't need 1M context. Earlier K2.x remains available but is generally superseded.
| Model | Context | Input $/1M | Output $/1M | Best at |
|---|---|---|---|---|
| kimi-k3 | 1M | $3.00 | $15.00 | Long-horizon agents, native vision, multi-file coding |
| kimi-k2.5 | 262K | $0.45 | $2.25 | Tool-use traces, agent loops at ~6.7x lower cost |
Official list prices verified 2026-08-23 (platform.kimi.ai; K2.5 via OpenRouter listing). See the full cheapest-Chinese-LLM price table for where Kimi sits against DeepSeek, GLM and Qwen.
Adoption signal: Kimi K3 is already running production workloads — see Harvey Tenet's switch to a Kimi K3 base and the K3 + GLM-5.2 open-stack teardown.
Python
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://api.chinamodelapi.com/v1")
resp = client.chat.completions.create(
model="kimi-k3",
messages=[
{"role": "system", "content": "You are a long-horizon coding agent. Cite file paths."},
{"role": "user", "content": f"REPO CONTEXT:\n{repo_blob}\n\nTASK: fix flaky integration tests."},
],
temperature=0.2,
)
print(resp.choices[0].message.content)Cost tip
K3 is frontier-priced relative to Flash models. Retrieve first, expand context only when confidence is low. Use K2.x / DeepSeek-V3 / GLM for the 80% short turns; reserve K3 for hard multi-file or multi-doc jobs.
FAQ
Is Kimi still on K2?
K2.x remains in the family, but the July 2026 flagship is Kimi K3. Product pages should lead with K3.
How big is K3?
Public coverage describes ~2.8T parameters and a 1M-token context window, with native vision.
Open weights?
Moonshot announced open weights for K3 (Modified MIT in some coverage) with weight publication around late July 2026. Until self-hosting is practical, use API access.
K3 vs GLM-5.2?
Both are Chinese frontier 2026 releases. K3 emphasizes 3T-class scale + long context + vision; GLM-5.2 emphasizes long-horizon coding with widely adopted open weights (its Aug 2026 successor GLM-5.3 doubles down on coding/agents). Run your eval harness.
What does Kimi cost per million tokens?
Official list (verified 2026-08-23): Kimi K3 is $3.00 input / $15.00 output per 1M tokens with 1M context; Kimi K2.5 is $0.45 / $2.25 with 262K context. A K2.5 agent loop costs roughly one-sixth of the same volume on K3.
K2.5 or K3 for agent loops?
Route tool-use traces and routine agent turns to K2.5, escalate the hard multi-file or multi-doc steps to K3. The OpenAI-compatible model id is the only thing that changes between them.