Moonshot has held the open-source size crown for 9 of the past 12 months (July 2025 – July 2026). K3 extends that streak into 3T-class territory — a psychological and technical barrier that makes open-weight models credible rivals to closed frontier models. The July 17 announcement moved real markets: US chip stocks (Nvidia, Micron) dipped on the news, as investors priced in a world where efficient open models reduce demand for frontier training compute. Demand was so strong that Moonshot tempor

| Model | Params | Context | Input / Output (per 1M) | Open weights | |---|---|---|---|---| | Kimi K3 | 2.8T | 1M | ¥20 / ¥100 | ✅ | | Kimi K2.6 | ~1T | 256K | ¥6.50 / ¥27 | ✅ | | Grok 4.5 | 1.5T (V9) | 500K | $2 / $6 | ❌ | | GPT-5.6 Sol | n/a | — | $5 / $25 | ❌ | | Claude Fable 5 | n/a | — | $10 / $50 | ❌ | Kimi K3 is priced as a premium flagship, but it's the only one of these you can download and self-host — and the first to prove 3T-class scale can be open.

Kimi K3: World's First Open-Source 3T-Class Model — Full Guide, Pricing & Impact

Q: Kimi K3 Architecture & Specs

| Spec | Kimi K3 | |---|---| | Parameters | 2.8T (world's first open-source 3T-class) | | Architecture | KDA (Kimi Delta Attention) + Attention Residuals | | MoE | 896 experts, 16 active (Stable LatentMoE) | | Context window | 1,048,576 tokens (1M) | | Multimodal | Native vision understanding | | Reasoning | Always-on, reasoning_effort: low / high / max | | Scaling efficiency | ~2.5x vs K2 | | Weights | Open-sourced by July 27, 2026 | The two architecture innovations worth knowing: - KDA (Kimi D

Q: Kimi K3 Pricing (API)

| Metric | Price | |---|---| | Input (cache miss) | ¥20.00 / 1M tokens | | Input (cache hit) | ¥2.00 / 1M tokens | | Output | ¥100.00 / 1M tokens | | Context window | 1M tokens | > 💡 Access note: Kimi K3 requires a recharge of at least ¥10 on the Moonshot platform. The ¥15 new-user voucher cannot be used on K3. Model ID: kimi-k3. Positioned as a premium flagship, K3's ¥100 output price puts it at the high end of the Chinese model market — roughly 3.7x Kimi K2.6's output price (¥27) — reflecting

⚡ Quick Answer: Kimi K3 is Moonshot AI's flagship model released July 17, 2026 — the world's first open-source 3T-class model at 2.8T parameters. It brings a 1M token context window, native vision understanding, and a new KDA (Kimi Delta Attention) architecture with 896-expert MoE. API pricing is ¥20 input / ¥100 output per 1M tokens (¥2 cached input). Weights were open-sourced by July 27. Its release briefly shook US chip stocks and set a new bar for open-weight models.

What Is Kimi K3?

Kimi K3 is the most capable model Moonshot AI (月之暗面) has ever built — and a milestone for open-source AI. With 2.8 trillion parameters, it's the first open-weight model to reach the 3T-class scale, a category previously reserved for closed frontier models.

The model is designed for three workloads:

Long-horizon coding — sustained engineering tasks with minimal supervision, large codebase understanding, and terminal tool use
End-to-end knowledge work — agentic workflows that span research, planning, and execution
Reasoning — always-on thinking mode with configurable reasoning effort

It also combines software engineering with visual reasoning — it can use screenshots and visual feedback for game dev, frontend, and CAD workflows.

Kimi K3 Architecture & Specs

Spec	Kimi K3
Parameters	2.8T (world's first open-source 3T-class)
Architecture	KDA (Kimi Delta Attention) + Attention Residuals
MoE	896 experts, 16 active (Stable LatentMoE)
Context window	1,048,576 tokens (1M)
Multimodal	Native vision understanding
Reasoning	Always-on, `reasoning_effort`: low / high / max
Scaling efficiency	~2.5x vs K2
Weights	Open-sourced by July 27, 2026

The two architecture innovations worth knowing:

KDA (Kimi Delta Attention) — designed to let information flow better across long sequences and deeper models
Attention Residuals (AttnRes) — complements KDA for training stability at scale

Combined with a sparser MoE (16 of 896 experts active), these changes give Kimi K3 roughly 2.5x better scaling efficiency than K2 — more capability per unit of compute.

Kimi K3 Pricing (API)

Metric	Price
Input (cache miss)	¥20.00 / 1M tokens
Input (cache hit)	¥2.00 / 1M tokens
Output	¥100.00 / 1M tokens
Context window	1M tokens

💡 Access note: Kimi K3 requires a recharge of at least ¥10 on the Moonshot platform. The ¥15 new-user voucher cannot be used on K3. Model ID: kimi-k3.

Positioned as a premium flagship, K3's ¥100 output price puts it at the high end of the Chinese model market — roughly 3.7x Kimi K2.6's output price (¥27) — reflecting its frontier-scale ambitions.

Why Kimi K3 Matters

1. The open-source scale record

Moonshot has held the open-source size crown for 9 of the past 12 months (July 2025 – July 2026). K3 extends that streak into 3T-class territory — a psychological and technical barrier that makes open-weight models credible rivals to closed frontier models.

2. Market impact

The July 17 announcement moved real markets: US chip stocks (Nvidia, Micron) dipped on the news, as investors priced in a world where efficient open models reduce demand for frontier training compute. Demand was so strong that Moonshot temporarily paused new K3 subscriptions under server load.

3. Competitive pressure on xAI

Elon Musk explicitly framed Grok 4.6 as the answer to Kimi K3 — a sign that even the most aggressive US labs now benchmark against Chinese open models. Kimi K3's arrival is a key reason the mid-2026 "open vs closed" frontier narrative shifted.

Kimi K3 vs The Field

Model	Params	Context	Input / Output (per 1M)	Open weights
Kimi K3	2.8T	1M	¥20 / ¥100	✅
Kimi K2.6	~1T	256K	¥6.50 / ¥27	✅
Grok 4.5	1.5T (V9)	500K	$2 / $6	❌
GPT-5.6 Sol	n/a	—	$5 / $25	❌
Claude Fable 5	n/a	—	$10 / $50	❌

Kimi K3 is priced as a premium flagship, but it's the only one of these you can download and self-host — and the first to prove 3T-class scale can be open.

How to Get Started with Kimi K3

Recharge at least ¥10 on platform.kimi.com (voucher not eligible)
Get an API key from the console
Call the API — OpenAI-compatible, base URL https://api.moonshot.cn/v1, model kimi-k3
Tune reasoning via top-level reasoning_effort: low / high / max (default max)

from openai import OpenAI

client = OpenAI(
    api_key="MOONSHOT_API_KEY",
    base_url="https://api.moonshot.cn/v1",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain KDA attention in one sentence."}],
)
print(resp.choices[0].message.content)

FAQ

Is Kimi K3 truly open source?

Yes. Moonshot confirmed full model weights would be released by July 27, 2026, with a technical report to follow. It's the first open-weight model at 3T-class scale (2.8T params).

How much does Kimi K3 cost?

API pricing is ¥20 per 1M input tokens (¥2 cached), ¥100 per 1M output tokens, with a 1M token context window. Requires a ¥10 minimum recharge.

Can I run Kimi K3 locally?

Technically yes once weights are released, but at 2.8T parameters it needs serious multi-node inference infrastructure — most teams will use the hosted API or quantized community builds.

What is KDA (Kimi Delta Attention)?

It's Kimi's new attention mechanism designed for better information flow across long sequences and deep models, paired with Attention Residuals for training stability at scale.

Does Kimi K3 support images?

Yes — native vision understanding is built in, including visual reasoning for coding, game dev, and CAD scenarios.

Bottom Line

Kimi K3 is not just another model release — it's the first time open-source has matched closed frontier scale. If you're building agentic coding or knowledge-work pipelines and value open weights, K3 deserves a serious look despite its premium pricing. If you want to watch the open-vs-closed race play out, it's the single most important release of the past month.

Related guides: