Kimi K3 Launch: 2.8T-Param Open-Source Flagship

​

Key Takeaways

• Largest open-source model ever: 2.8 trillion total parameters (MoE), the first open model in the 3-trillion-parameter class; full weights promised by July 27, 2026 (UTC+8)
• Kimi Delta Attention + 1M context: an exact 1,048,576-token window with flat pricing across the whole range — no context-tier markups
• #1 on Frontend Code Arena: jumped from Kimi K2.6’s #18 to first place on LMArena’s frontend leaderboard, winning 6 of 7 frontend domains
• Strong reasoning/agent scores: GPQA Diamond 93.5%, Terminal-Bench 2.1 88.3%, BrowseComp 91.2%, Humanity’s Last Exam (with tools) 56.0%
• Native multimodal + always-on thinking: native image/video input; reasoning runs at max effort by default (reasoning_effort="max")
• Same price on APIYI: listed at $3.00/$15.00 per 1M tokens (input/output), identical to Moonshot’s official pricing, with cache-hit input as low as $0.30

​

Background

​

Deep Dive

​

Core Features

​

Benchmark Highlights

​

Technical Specs

​

Practical Usage

​

Recommended Scenarios

​

Code Example

from openai import OpenAI

client = OpenAI(
    api_key="your-apiyi-key",
    base_url="https://api.apiyi.com/v1"
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "system", "content": "You are a rigorous senior frontend engineer."},
        {"role": "user", "content": "Build a kanban board component with drag-and-drop sorting in React + Tailwind"}
    ]
)
print(response.choices[0].message.content)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,<base64>"}},
            {"type": "text", "text": "Recreate this design mockup in HTML + CSS"}
        ]
    }]
)

​

Best Practices

• Skip sampling parameters: temperature and friends are fixed by Moonshot — omit them from requests
• Budget for output tokens: thinking always runs at max effort, so output usage is higher than typical models; max_completion_tokens defaults to 131,072 and can be raised as needed
• Leverage caching: Kimi K3 caches context automatically; cache-hit input bills at $0.30/1M (1/10 of the miss price), a big win for multi-turn chats and repeated prefixes
• Feed long material directly: with flat pricing up to 1M tokens, whole codebases and long documents can go in without tier-crossing cost surprises

​

Pricing & Availability

​

Pricing

​

Stack Recharge Bonuses

• Recharge $100, get 10% bonus → an easy ~9% effective discount
• Larger tiers reach up to ~17% off (tier-dependent; see the recharge promotions FAQ)

​

Summary & Recommendations

1. The open-source flagship answer: 2.8T parameters with weights landing a week later — “open catching up to closed” has never had this much weight behind it
2. Frontend and agents are the headline strengths: #1 on Frontend Code Arena and 88.3% on Terminal-Bench 2.1 make it a first pick for frontend work and long-horizon agents
3. Friendly price structure: $3/$15 is a value tier for a 1M-context model, and 1/10-price cache hits plus flat context pricing keep long-context costs under control