Here is the part that matters for you: Qubrid AI is a day-0 launch partner. The Kimi K3 API is live on our platform right now. No waitlist, no queue. If you have a Qubrid account, you are one model ID away from building on the biggest open model in the world.

In this post, we break down what Kimi K3 actually is, what it costs, how it compares to the Kimi K2 family, and how to make your first K3 API call in under five minutes.

What Is Kimi K3?

Kimi K3 is the new flagship model from Moonshot AI, the team behind the Kimi K2 series (K2, K2.5, K2.6, and K2.7 Code). It is built on two architectural ideas that make its scale practical to serve: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals. Together they let a 2.8 trillion parameter model run a full million-token context without the memory blowup that makes most long-context models impractical in production.

The quick spec sheet:

For context on how fast this space is moving: DeepSeek V4-Pro, the previous largest open release, sits at 1.6 trillion parameters. K3 clears it by more than a trillion.

What the Kimi K3 API Is Built For

Moonshot positions K3 around three workloads, and early testing backs it up:

Software engineering at repository scale. K3 is strong at navigating large codebases, calling tools, debugging, and iterating against tests, logs, and runtime feedback. With 1M tokens of context, an entire mid-sized monorepo fits in a single request. Your coding agent stops juggling context and starts actually reading the code.

Long-horizon agent workflows. Thinking mode is always on, with a reasoning_effort parameter to control depth. Multi-step agents keep the full trace in context, so decision quality on step 200 is as informed as step 2.

Knowledge work and deep reasoning. Contract analysis, research synthesis, compliance review, any task where the answer depends on reading a lot of material carefully rather than retrieving fragments of it.

It is also natively multimodal. K3 accepts images and video alongside text, and early community tests show surprisingly strong results on visual reasoning, 3D generation, and game code that actually runs.

Kimi K3 Benchmarks: The Open vs Closed Gap Just Closed

The number everyone is talking about: Kimi K3 launches at 57 on the Artificial Analysis Intelligence Index. For perspective on what that means:

That makes K3 the highest-scoring open-weight model ever evaluated by Artificial Analysis, a full 6 points above GLM-5.2, the previous open-weights leader. More striking: it edges past Claude Opus 4.8 and sits within 3 points of the top closed frontier models. The long-standing assumption that open models trail closed ones by six to twelve months does not survive this launch.

Moonshot's own reported numbers reinforce the picture:

• 91.2% on BrowseComp in a single-agent setup, with no context compression or additional context-management techniques. This is where the 1M window shows up as capability, not just capacity: the agent simply keeps everything in context.

• 1687 Elo on GDPval-AA v2, an evaluation of real economically valuable work spanning 44 occupations across nine industries.

• 1527 Elo on AA-Briefcase, ranking second in the provider-published comparison.

Early independent testing echoes the trend on coding: community-run evaluations place K3 just behind Claude Fable 5 on DeepSWE-style software engineering tasks and ahead of Opus 4.8 on terminal-based agentic benchmarks. As always, provider-published and day-one community numbers should be read as directional until the full independent leaderboards (SWE-bench Verified, Terminal-Bench, LMArena) fill in over the coming weeks. But the direction is unambiguous.

The Research Behind Kimi K3: Why 1M Context Actually Works

Plenty of models advertise long context. What makes K3 different is that its architecture was redesigned to make million-token inference economically viable, and that research story is worth understanding if you are going to build on it.

Kimi Delta Attention (KDA). Standard transformer attention scales quadratically: doubling context roughly quadruples compute, and the KV cache grows linearly with every token you keep in memory. At 1M tokens, that is what makes most long-context models slow and expensive to serve. KDA is a hybrid linear attention mechanism that dramatically reduces KV cache memory and boosts decoding throughput at long context lengths, which is how K3 serves the full 1M window at flat pricing with no long-context surcharge.

Attention Residuals. The historical weakness of linear attention is quality degradation, models get faster but lose precision at recalling details from deep in the context. Attention Residuals are Moonshot's answer: they preserve the information flow that pure linear attention loses, so the model stays accurate at position 900,000 in the context, not just position 9,000. The BrowseComp result above, achieved without any context compression tricks, is the practical evidence that this works.

Scale as a research bet. At 2.8 trillion total parameters in a Mixture-of-Experts configuration, K3 is a deliberate jump rather than an increment. The open-weight field spent the past year clustered between 500B and 1T parameters. Moonshot's bet is that sparse MoE routing lets you scale total capacity aggressively while keeping per-token active compute, and therefore inference cost, in a usable range. The AA Intelligence Index result suggests the bet paid off.

For builders, the takeaway is concrete: K3's long context is not a marketing number with an asterisk. The architecture was built for it, the pricing reflects it, and the benchmarks validate it.

Kimi K3 Pricing

K3 uses flat pay-as-you-go pricing with no tiering by context length. The full 1M window is included at the standard rate:

The cache-hit rate is the number to pay attention to. Context caching is automatic on K3, with no cache IDs or TTL management required. Keep your long prefix stable (system prompt, repo context, knowledge base) and repeated requests bill input at $0.30 per million instead of $3.00. For coding agents that resend large context every turn, that is a 90% cut on the bulk of your bill.

Reasoning tokens are billed as output tokens, so there is no separate thinking SKU to budget for.

Kimi K3 vs Kimi K2.6 and K2.7 Code

The practical read: K2.7 Code is still an excellent, cost-efficient workhorse for coding agents that fit in 256K context. Reach for K3 when you need the million-token window, video input, or maximum reasoning quality on genuinely hard problems. Both run on Qubrid behind the same API, so switching between them is a one-line change.

How to Use the Kimi K3 API on Qubrid AI

The Qubrid inference API is OpenAI-compatible, so if you have ever written code against the OpenAI SDK, you already know how to use K3.

Step 1: Get your API key. Sign up on the Qubrid AI platform and generate a key from your dashboard.

Step 2: Make your first call.

python

from openai import OpenAI

# Initialize the OpenAI client with Qubrid base URL
client = OpenAI(
    base_url="https://platform.qubrid.com/v1",
    api_key="QUBRID_API_KEY",
)

response = client.chat.completions.create(
    # Must match the exact model ID from the docs — variations will cause errors.
    model="moonshotai/Kimi-K3",
    messages=[
      {
        "role": "user",
        "content": "Explain the main benefits of using a chat completion API for text generation."
      }
    ],
    max_tokens=32768,
    temperature=1,
    top_p=0.95,
    stream=False,
    frequency_penalty=0,
    presence_penalty=0
)

print(response.choices[0].message.content)

Step 3: Turn on the good stuff. K3 supports streaming with separate reasoning and answer deltas, strict JSON schema structured output, function calling with tool_choice control, and vision input via base64-encoded images. All of it works through the same endpoint.

A few implementation notes worth knowing on day 0:

• reasoning_effort currently supports the max level, with more levels coming. Thinking is always enabled, so do not pass the older K2-style thinking parameter.

• Sampling parameters like temperature and top_p are fixed on K3. Omit them from requests.

• In multi-turn and tool-calling flows, pass the complete assistant message back, not just the content field.

• To maximize cache hits, keep your system prompt and long context identical across requests.

Why Developers Run Kimi Models on Qubrid AI

Day-0 access is not a one-off for us. Qubrid brought GLM-5.2, MiniMax M3, and the Qwen3-Coder series to production APIs on launch day, and we host the full Kimi lineup: K2.5, K2.6, K2.7 Code, and now K3, all behind one API key.

That matters because real applications route by task. Cheap classification calls go to a small model, standard coding agents run on K2.7 Code, and the hardest reasoning and repo-scale work goes to K3. One platform, one key, one integration.

Add OpenAI-compatible endpoints, transparent token-based pricing, and infrastructure built for long-context inference at production throughput, and the path from "K3 launched last night" to "K3 is in our product" is a single afternoon.

Frequently Asked Questions

What is the Kimi K3 API? The Kimi K3 API provides programmatic access to Moonshot AI's Kimi K3 model for chat, coding, agentic, and multimodal workloads. On Qubrid AI, it is available from day 0 through an OpenAI-compatible endpoint.

How do I get Kimi K3 API access? Create a Qubrid AI account, generate an API key, and call the kimi-k3 model ID. Access is live now with no waitlist.

How much does the Kimi K3 API cost? $3.00 per million input tokens, $0.30 per million for cached input, and $15.00 per million output tokens, with the full 1M context included at that flat rate.

What is the Kimi K3 context window? 1,048,576 tokens (1M), with output configurable up to 1,048,576 tokens as well.

Is Kimi K3 open source? Yes, K3 is an open-weight release, continuing the pattern of the Kimi K2 family. If you would rather not self-host a 2.8T parameter model, the hosted K3 API on Qubrid is the fastest path.

How does Kimi K3 score on Artificial Analysis? Kimi K3 launches at 57 on the Artificial Analysis Intelligence Index, the highest score ever recorded by an open-weight model. It places above Claude Opus 4.8 (56) and within 3 points of the index leader, Claude Fable 5 (60).

Is Kimi K3 the best open-source model? By the Artificial Analysis Intelligence Index, yes. At 57, K3 leads all open-weight models, 6 points clear of the previous leader, GLM-5.2 (51). It is also the largest open model released at 2.8 trillion parameters.

Is Kimi K3 better than K2.6 or K2.7 Code? K3 is a generational jump: 2.8T parameters vs roughly 1T, 1M context vs 256K, video input, and always-on reasoning. K2.6 and K2.7 Code remain the cost-efficient choice for workloads that fit in 256K context.

Does the Kimi K3 API support function calling and structured output? Yes. K3 supports function calling with tool_choice control and strict JSON schema structured output, making it well suited for production agents.

Can I use my existing OpenAI SDK code? Yes. Qubrid's endpoint is OpenAI-compatible. Point base_url at Qubrid, set the model to kimi-k3, and your existing code runs.

Start Building with Kimi K3 Today

The largest open model ever released, a million tokens of context, and frontier-level coding, available through a production API on the day it launched. That is the whole pitch.

Get your API key on Qubrid AI and make your first Kimi K3 call today.

https://platform.qubrid.com/model/kimi-k3