Kimi K3 Review: The 2.8-Trillion-Parameter Open-Source Model That Challenges Claude and GPT
On the night of July 16, 2026, Moonshot AI quietly flipped a switch: a banner appeared atop the Kimi API docs — "🎉 Kimi K3 is live!" No press conference. Just a tech blog, pricing page, and a model ID you could call immediately. This guide for developers and model buyers covers KDA architecture, the full benchmark picture, pricing, four access paths, the July 27 open-weight release, a five-step Runbook, and a scenario matrix.
Table of Contents
Pain Points: Why K3 Forces a Model Routing Reset
- The open/closed intelligence gap is shrinking. K3 scores 57.1 on Artificial Analysis Intelligence Index v4.1 (4th place) — just 2.8 points behind Claude Fable 5 (59.9). Open weights are now in the same conversation as frontier closed APIs.
- Long context vs real bills. Competitors cap at 200K–400K with length surcharges. K3 offers 1M tokens at flat $3/$15 pricing, with 90%+ cache hit rates in coding — effective input as low as $0.30/M.
- Single-vendor policy risk. The Claude Fable 5 export-control shutdown showed production agents on one closed API can go dark in 90 minutes. K3's July 27 full weight release adds a self-hosting escape hatch.
What Is Kimi K3?
Kimi K3 is a 2.8-trillion-parameter MoE model from Moonshot AI — the world's first open 3T-class system, surpassing DeepSeek V4 Pro (1.6T) by nearly 75%.
| Spec | Detail |
|---|---|
| Total Parameters | 2.8 trillion |
| Architecture | KDA + AttnRes + Stable LatentMoE |
| Active Experts | 16 of 896 (1.8% sparsity) |
| Context Window | 1,048,576 tokens (1M) |
| Input Modalities | Text, image, video |
| API Model ID | kimi-k3 |
| Open Weights | July 27, 2026 |
Only 16 of 896 experts activate per forward pass. Native vision plus a 1M-token window targets long-horizon coding, document reasoning, and knowledge work. One-liner: an open, vision-capable, long-memory coding AI priced ~40% below Claude Opus 4.8, with full weights coming July 27.
Why This Release Matters
The last 18 months were rough for Moonshot as DeepSeek eroded market share. K3 is a striking comeback:
- For 9 of the past 12 months, Kimi models held the largest open-source parameter record;
- Launch timed for the eve of WAIC 2026 (July 17–20, Shanghai);
- ARR crossed $300M by June 2026; 6th funding round at $31.5B pre-money valuation;
- API revenue is 70%+ of total; overseas paid users up 400%.
This is a fast-growing business making a serious technical statement — not a vanity scale play.
The Architecture: Three Genuine Innovations
1. Kimi Delta Attention (KDA)
Full attention scales quadratically — at 1M tokens, KV cache memory becomes catastrophic. KDA alternates 3 linear-attention layers : 1 full-attention layer:
- 75% less KV cache memory;
- Up to 6.3× faster decoding at 1M contexts;
- Matches or beats full-attention baselines on short, long, and RL-extended tasks — no capability tradeoff.
2. Attention Residuals (AttnRes)
Selective retrieval across depth pulls high-value early-layer representations instead of uniform accumulation — ~25% higher training efficiency at under 2% extra compute.
3. Stable LatentMoE
| Technique | Role |
|---|---|
| Quantile Balancing | Expert allocation from router-score quantiles — no fragile heuristics |
| Per-Head Muon | Per-head optimization for adaptive large-scale training |
| SiTU | Improved activation control |
| Gated MLA | Better attention selectivity |
Net result: ~2.5× better scaling efficiency vs Kimi K2 on the same compute budget.
Benchmark Results: Where It Wins and Where It Doesn't
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | GLM-5.2 |
|---|---|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | 46.2 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 63.7 |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 67.3 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 13.0 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | — |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 | 12.9 |
| GPQA-Diamond | 93.5 | 92.6 | 94.1 | 91.0 | 91.2 |
| MMMU-Pro | 81.6 | 81.2 | 83.0 | 78.9 | — |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 | — |
| HLE-Full | 43.5 | 53.3 | 44.5 | — | — |
SWE Marathon (sustained multi-hour coding) is K3's headline win at 42.0 — a 7-point gap over Fable 5. OmniDocBench leadership (91.1) reflects vision + 1M context synergy. Overall index: K3 at 57.1 (#4). Caveat: Moonshot self-reported; harnesses differ (Kimi Code vs Codex vs Claude Code).
Pricing: How Does It Stack Up?
| Model | Input $/1M | Output $/1M | Cache-Hit Input | Context |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 | 1M |
| Claude Sonnet 5 | $3.00 | $15.00 | — | 200K |
| Claude Opus 4.8 | $5.00 | $25.00 | — | 200K |
| GPT-5.5 | $5.00 | $30.00 | — | 400K |
| DeepSeek V4 Pro | $1.74 | $3.48 | $0.145 | 128K |
K3 matches Sonnet 5 standard pricing but delivers 5× context. Mooncake split-inference drives 90%+ cache hits in Kimi Code — effective average input ~$0.55/M (OpenRouter 7-day weighted average). vs Opus 4.8: stronger on several benchmarks at 60% input / 40% output cost.
How to Use Kimi K3 Right Now
Option 1: Chat (no setup)
kimi.com — sign up with Google. K3 runs at max reasoning effort. No credit card.
Option 2: API
API key at platform.kimi.ai.
Option 3: OpenRouter
Model ID: moonshotai/kimi-k3 — official $3/$15, no markup, full 1M context.
Option 4: Wait for open weights (July 27)
Full weights on Hugging Face. Production needs a 64+ accelerator supernode. Trained with MXFP4 weights / MXFP8 activations; Day-0 support expected in transformers, vLLM, SGLang.
Kimi K3 vs. The Competition
| Use Case | Best Pick | Why |
|---|---|---|
| Long sustained coding sessions | Kimi K3 | Leads SWE Marathon; 1M context prevents mid-task loss |
| Complex repo-level bug fixes | Claude Fable 5 | FrontierSWE / SWE-bench Pro lead |
| Terminal/tool-heavy agents | GPT-5.6 Sol | Terminal Bench + Coding Agent Index |
| Multimodal document analysis | Kimi K3 | Best OmniDocBench; vision + 1M context |
| Cost-sensitive production | DeepSeek V4 Pro | $3.48/M output, far cheaper |
| Open-source self-hosting (post 7/27) | Kimi K3 | Largest open weights available |
| Deepest reasoning (HLE-Full) | Claude Fable 5 | 53.3 vs 43.5 — wide margin |
The Open-Source Promise: July 27
Moonshot committed to full weight release July 27, 2026 (Modified MIT). K3 becomes:
- Largest downloadable open-source model ever;
- First open model above 2 trillion parameters;
- New fine-tuning and research foundation for 2026.
Citable Technical Facts
- Scale: 2.8T params — 75% above DeepSeek V4 Pro (1.6T).
- Sparsity: 16/896 experts active (1.8%).
- KDA: 75% KV cache reduction; 6.3× decode speedup at 1M tokens.
- Cache economics: 90%+ hit rate → ~$0.55/M effective input.
- Intelligence index: 57.1 — 2.8 points from #1.
Five-Step API Runbook
Step 2 Pick path: free web trial / official API / OpenRouter moonshotai/kimi-k3
Step 3 Configure OpenAI SDK: base_url=https://api.moonshot.ai/v1, model=kimi-k3
Step 4 Pilot 10-20 SWE Marathon-style tasks; log quality, tokens, cache hits
Step 5 Hybrid routing: long code/docs → K3; repo bugs → Fable 5; terminal agents → GPT-5.6 Sol
Frequently Asked Questions
Is Kimi K3 available for free?
Yes on kimi.com. API is pay-per-token at $3/$15 per 1M.
Can I run it locally?
Weights July 27. Need 64+ GPUs for production — not a laptop LLM.
K3 vs DeepSeek V4 Pro?
K3 wins on scale, context, benchmarks; DeepSeek wins on cost ($3.48/M output).
Is 1M context useful?
Yes for full repos, long docs, long-memory agents — flat pricing encourages using it.
Low/high reasoning modes?
Coming in subsequent updates; only max at launch.
Trust the benchmarks?
Self-reported with vendor harnesses — validate with your own pilots.
Bottom Line
Kimi K3 is the most capable open-source AI model ever shipped to API. It doesn't win every benchmark — Fable 5 and GPT-5.6 Sol still lead specific tasks — but it's competitive across the board, dominates long-horizon coding and document understanding, and ships 1M context at Sonnet-class pricing. The July 27 weight drop is the story to watch.
Dates to bookmark: Now → try at kimi.com · July 17–20 → WAIC Shanghai · July 27 → Hugging Face weights.
Running K3 agents on a laptop or generic Linux VPS means sleep-interrupted long loops, commingled API keys and production repos, and no co-located Apple toolchain (Xcode, Fastlane, notarytool). For 7×24 unattended agents — Kimi Code plus iOS CI or an OpenClaw gateway — renting a VPSMAC M4 Mac cloud node (native macOS, SSH + launchd, isolated from your daily machine) is usually the more stable production path than personal hardware or Linux VPS.
Sources: Moonshot official blog · Kimi API docs · Artificial Analysis · OpenRouter pricing
Data as of 2026-07-16. Benchmarks are Moonshot self-reported. Verify latest official docs before production decisions.