Kimi K3 Review: The 2.8-Trillion-Parameter Open-Source Model
That Challenges Claude and GPT

TL;DR: Moonshot AI just released Kimi K3 — the world's largest open-source AI model at 2.8 trillion parameters. It has a 1M token context window, native vision, beats Claude Fable 5 and GPT-5.6 Sol on several coding benchmarks, and costs $3/$15 per million tokens. Full weights drop July 27. Here's everything you need to know.

On the night of July 16, 2026, Moonshot AI quietly flipped a switch. A banner appeared at the top of the Kimi API docs: "Kimi K3 is live!" No press conference. No countdown timer. Just a technical blog, a pricing page, and a model ID you could call immediately. This article is for AI developers, researchers, and product engineers who need a structured answer to the questions that matter most: what Kimi K3 is, how its architecture works, how it compares to Claude Fable 5 and GPT-5.6 Sol, what the API costs, how to integrate it today, and why the July 27 open-weight release changes the open-source landscape.

01 What Is Kimi K3? Why This Quiet Launch Matters

Kimi K3 is the world's largest open-source AI model by parameter count — 2.8 trillion (2.8T), surpassing the previous record holder DeepSeek V4 Pro (1.6T) by nearly 75%. It uses a sparse mixture-of-experts (MoE) architecture, activating 16 out of 896 experts per forward pass, paired with a 1 million token context window and native vision. It is built for complex coding, long-document reasoning, and knowledge work.

Bottom line: Kimi K3 is an open-source, vision-capable coding model with an unusually long memory, priced below Claude Opus 4.8 on output, with full weights scheduled for public release on July 27, 2026.

The model is already live on kimi.com, the Kimi mobile app, Kimi Code, and the Moonshot API. You can try it now with a free Google account — no credit card required.

Moonshot AI spent the last 18 months under pressure from DeepSeek's rise. K3 is a direct comeback — but engineers should understand the real constraints before switching production workloads:

• Closed API dependency until July 27: Until weights ship, production use depends entirely on Moonshot's cloud. Quotas, latency, and pricing can change without notice.
• 1M context is not free compute: Flat token pricing helps, but long Agent runs still need persistent dev environments and reliable toolchains — context alone does not solve orchestration.
• Self-hosting is not realistic for most teams: Production inference requires a supernode with 64+ accelerators. Laptops and single-GPU boxes are out of scope.
• Benchmarks are self-reported: Moonshot used Kimi Code; GPT used Codex; Claude used Claude Code. Independent reproduction is still in progress — treat scores as directional.

The strategic context behind the launch is equally important:

• For 9 of the past 12 months, Kimi models held the record for the largest open-source model by parameter count;
• The release was timed for the eve of the 2026 World AI Conference (WAIC) in Shanghai (July 17–20);
• As of June 2026, Moonshot AI's ARR crossed $300M, with a $31.5B pre-money valuation after its 6th funding round this year;
• API revenue now makes up over 70% of total revenue, with overseas paid users growing 400%.

This is not a vanity-scale project from a struggling lab. It is a fast-growing business making a serious technical statement on the eve of China's largest AI conference.

02 Architecture: KDA, AttnRes, and Stable LatentMoE

Most "biggest model" announcements add parameters without fixing scaling bottlenecks. Kimi K3 introduces three architectural innovations aimed at long-context efficiency and stable extreme-sparsity training — a different engineering bet from the models covered in our OpenRouter June model rankings analysis.

Kimi Delta Attention (KDA): Rethinking How Transformers Remember

Standard full attention scales quadratically with context length. At 1 million tokens, KV cache memory consumption alone becomes catastrophic. KDA is a hybrid linear attention mechanism:

• Alternates linear-attention and full-attention layers in a 3:1 ratio — three cheap linear layers handle local structure, one full-attention layer preserves global flow;
• Reduces KV cache memory by up to 75%;
• Delivers up to 6.3× faster decoding at 1M-token contexts;
• Matches or beats full-attention baselines on short-context, long-context, and reinforcement learning tasks — no obvious capability tradeoff.

Think of full attention as memorizing every detail at once. KDA behaves more like an efficient index: fast lookups most of the time, precise recall when it matters. That is why K3 can offer a genuine 1M-token window at flat pricing rather than a theoretical maximum nobody uses.

Attention Residuals (AttnRes): Fixing Information Loss Across Depth

• Standard residual connections accumulate representations uniformly across depth, diluting early-layer signals in deeper layers;
• AttnRes adds selective retrieval across depth — the model can pull high-value representations from earlier layers directly;
• Moonshot reports approximately 25% higher training efficiency at under 2% additional compute cost.

Stable LatentMoE: Making Extreme Sparsity Work

Kimi K3 has 896 experts and activates 16 per forward pass (1.8% sparsity). At this level, naive routing breaks down. Supporting techniques:

03 Benchmarks: Kimi K3 vs Claude Fable 5 vs GPT-5.6 Sol

Below are Moonshot's self-reported core benchmarks. Read them alongside our Claude Fable 5 engineering guide if your primary workload is repo-scale bug fixing rather than multi-hour coding sessions.

SWE Marathon is the headline number. It measures sustained, long-horizon coding work — the benchmark closest to "writing code for hours." Kimi K3 leads at 42.0, a 3-point gap over GPT-5.6 Sol and 7 points over Claude Fable 5. If your use case involves extended coding sessions without context resets, this is the score to weight most heavily.

On OmniDocBench (mixed text, tables, charts, and images), Kimi K3 leads all listed models — consistent with native vision plus 1M-token context working together. On the Artificial Analysis Intelligence Index v4.1, K3 scores 57.1, placing 4th overall behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9) — a gap of just 2.8 points to the leader for an open-weight-class model competing with closed frontier systems.

Important caveat: these are Moonshot self-reported benchmarks using different harnesses per vendor. Treat them as directionally useful, not definitive, until independent reproduction lands.

04 Kimi K3 API Pricing and 6-Step Integration Guide

K3 matches Claude Sonnet 5's standard rate ($3/$15) but offers a 5× larger context window. The cache-hit price drops to $0.30/M. Moonshot reports 90%+ cache hit rates in Kimi Code workflows thanks to Mooncake split-inference architecture, which pushes effective average input cost to roughly $0.55/M tokens in real usage — a figure OpenRouter's 7-day weighted average corroborates empirically. Compared to Claude Opus 4.8, K3 beats it on several benchmarks at 60% of the input cost and 40% of the output cost.

6 steps to integrate Kimi K3 via API:

1. Create a Moonshot account: Sign up at platform.kimi.ai or kimi.com (Google login supported).
2. Generate an API key: Create a key in the console and store it securely — it is shown only once.
3. Confirm the model ID: Use kimi-k3; on OpenRouter, use moonshotai/kimi-k3 at official pricing with no markup.
4. Configure an OpenAI-compatible client: Set base_url to https://api.moonshot.ai/v1.
5. Send a smoke-test request: Run a short prompt to verify connectivity, latency, and billing or free-tier status.
6. Enable cache-aware workflows: Reuse system prompts and tool definitions in Kimi Code or Agent pipelines to capture 90%+ cache hits and lower effective input cost.

from openai import OpenAI

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

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Analyze this codebase and identify performance bottlenecks..."}]
)
print(response.choices[0].message.content)

05 Which Model Should You Use? July 27 Open Weights

Moonshot has committed to releasing full model weights on July 27, 2026 via Hugging Face. When they drop, Kimi K3 becomes:

• The largest downloadable open-source model ever released;
• The first open model above 2 trillion parameters;
• A new foundation for open-source fine-tuning and research — with day-0 support expected in transformers, vLLM, and SGLang, plus MXFP4/NVFP4 quantized variants.

The model was trained with MXFP4 weights and MXFP8 activations — quantization-aware from the start, designed for efficient inference on modern hardware.

Dates to bookmark: July 17–20 (WAIC Shanghai, more announcements expected) → July 27 (full K3 weights on Hugging Face).

06 Citeable Technical Data and Sources

• Total parameters: 2.8 trillion (2.8T), largest open-source model released to date
• MoE configuration: 896 experts, 16 active per forward pass (1.8% sparsity)
• Context window: 1,048,576 tokens (1M)
• KDA efficiency: KV cache −75%, up to 6.3× faster decoding at 1M tokens
• AttnRes: ~25% training efficiency gain at <2% extra compute
• Scaling vs K2: ~2.5× better scaling efficiency
• API pricing: $3 / $15 per 1M tokens (input/output); cache-hit $0.30/M; effective ~$0.55/M in cached Agent workflows
• ARR (June 2026): crossed $300M; pre-money valuation $31.5B
• Artificial Analysis Index v4.1: K3 scores 57.1, 4th globally
• Open weights date: July 27, 2026 (Hugging Face)

Primary references below. Capabilities, pricing, and release schedules may change — verify against official sources after publish:

Official sources:

Moonshot AI — Kimi K3 Official Technical Blog

Kimi API Platform — Documentation and Pricing

Third-party analysis and coverage:

Artificial Analysis — Intelligence Index v4.1 Rankings

OpenRouter — moonshotai/kimi-k3 Pricing and Usage

Kimi K3 does not win every benchmark — Claude Fable 5 and GPT-5.6 Sol still lead specific coding and reasoning tests — but it is competitive across the board, leads on long-horizon coding and document understanding, and ships a 1M-token context window that closed competitors cannot match at this price. For engineering teams, API-only workflows still carry network latency, quota risk, and toolchain integration overhead; 2.8T self-hosting remains impractical before July 27 and requires supernode-scale hardware afterward; and shared virtual Macs or cloud VMs running Xcode, Metal, and Agent toolchains introduce hypervisor overhead and compatibility friction that degrade long-running Agent stability. If your production environment needs zero-overhead Apple Silicon, 24/7 uptime, reliable iOS CI/CD, and AI Agent automation — for example, running Kimi Code or Claude Code on a persistent physical Mac while routing to multiple model APIs — ZUKCLOUD bare metal Mac mini cloud nodes are typically the more controlled option: dedicated physical hardware, no virtualization tax, elastic daily/weekly/monthly billing. See pricing and order, or read the bare metal architecture manifesto for Agent-grade hosting logic.

07 Frequently Asked Questions

Q: Is Kimi K3 available for free?
A: Yes — on kimi.com with a free account. K3 runs at max reasoning effort by default. API usage requires pay-per-token billing ($3/$15 per 1M tokens).

Q: Can I run Kimi K3 locally?
A: Not yet. Full weights release on July 27, 2026 via Hugging Face. Production inference requires a supernode with 64+ accelerators — not a laptop or single-GPU setup.

Q: How does Kimi K3 compare to DeepSeek V4 Pro?
A: K3 has nearly double the parameters (2.8T vs 1.6T), 5× the context (1M vs 128K), and stronger scores on several coding benchmarks — but DeepSeek output costs $3.48/M vs K3's $15/M. Cost-sensitive workloads may still favor DeepSeek. See our DeepSeek compute ecosystem analysis for broader context.

Q: Is the 1M token context window actually useful?
A: Yes — for whole-codebase analysis, end-to-end legal or research documents, and multi-session Agent memory. Flat pricing with no length surcharge, combined with high cache hit rates, keeps real costs manageable.

Q: When are low/high reasoning effort modes coming?
A: Moonshot says low and high modes arrive in subsequent updates. Only max is available at launch.