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

Real pain points before you pick a model:

• Parameter count alone misleads: MoE models activate a tiny fraction of experts per token — check active experts and context efficiency.
• Benchmark harnesses differ: Moonshot uses Kimi Code; GPT uses Codex; Claude uses Claude Code — treat cross-vendor scores as directional.
• 1M context without architecture support is marketing: KV cache cost explodes without mechanisms like KDA.
• Waiting for weights delays production: The API is live now; self-hosting can wait until July 27.

Kimi K3 is a 2.8T-parameter MoE model from Beijing-based Moonshot AI — the first open model in the 3T class, surpassing DeepSeek V4 Pro (1.6T) by nearly 75%. It activates 16 of 896 experts per forward pass, ships a genuine 1,048,576-token context, and accepts text, image, and video input. Model ID: kimi-k3. Reasoning runs at max effort at launch.

On the night of July 16, Moonshot flipped a switch — a banner on the API docs, no press conference. The quiet launch contrasts with the scale: the biggest open-source AI model ever built.

Why this release matters: DeepSeek's rise eroded Moonshot's position over 18 months. K3 is a comeback — Kimi models held the open-source size record for 9 of the past 12 months; ARR crossed $300M by June 2026; a 6th funding round landed at $31.5B pre-money; API revenue is 70%+ of total with overseas paid users up 400%. Timing aligns with the eve of WAIC 2026 in Shanghai. This is a fast-growing business making a technical statement, not a vanity scale play.

Most "biggest model" launches add compute. K3 introduces three engineering fixes for long context and extreme MoE sparsity.

Kimi Delta Attention (KDA) alternates cheap linear-attention layers with full-attention layers in a 3:1 ratio. Result: up to 75% less KV cache memory and up to 6.3× faster decoding at 1M-token contexts, matching or beating full-attention baselines on short, long, and RL tasks — no capability tradeoff. That is why the 1M window is priced flat, not theoretical.

Attention Residuals (AttnRes) replace uniform depth accumulation with selective retrieval across layers — roughly 25% higher training efficiency at under 2% extra compute.

Stable LatentMoE makes 1.8% sparsity (16/896 experts) trainable via Quantile Balancing, Per-Head Muon, SiTU activation, and Gated MLA — enabling ~2.5× better scaling efficiency vs Kimi K2. Training used MXFP4 weights and MXFP8 activations for quantization-aware deployment.

Moonshot-published scores below. Independent reproduction is ongoing — treat as directional.

SWE Marathon — sustained long-horizon coding — is the headline: K3 leads at 42.0, a 7-point gap over Claude Fable 5. OmniDocBench leadership fits native vision plus 1M context. On Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 (4th overall), just 2.8 points behind #1 Fable 5 (59.9) and ahead of most open weights by a wide margin.

K3 matches Claude Sonnet 5 standard pricing but offers 5× context. Mooncake split-inference drives 90%+ cache hit rates in Kimi Code workflows — effective average input cost can fall to roughly $0.55/M (OpenRouter 7-day weighted average). vs Opus 4.8: stronger on several benchmarks at ~60% input and ~40% output cost.

• Chat free: kimi.com — Google sign-in, max reasoning, no credit card.
• Official API: OpenAI-compatible at https://api.moonshot.ai/v1, key from platform.kimi.ai.
• OpenRouter: moonshotai/kimi-k3 — official pricing, no markup, full 1M context.
• Self-host July 27: Full weights on Hugging Face; production needs 64+ accelerators, not a laptop.

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 for bottlenecks..."}]
)
print(response.choices[0].message.content)

Primary official sources — re-open links after upstream updates:

https://kimi.com/blog/kimi-k3

https://platform.kimi.ai

Six-step production rollout:

1. Create API credentials: Register at platform.kimi.ai, add billing, generate an API key.
2. Pick a route: Direct Moonshot API or OpenRouter moonshotai/kimi-k3 inside your existing OpenAI SDK stack.
3. Enable prompt caching: Configure cache keys for repo-scale and agent loops to capture 90%+ hit rates.
4. Plan 1M context usage: Whole-codebase analysis, legal/research docs, multi-session agent memory — flat pricing makes full-window use practical.
5. Hybrid routing: Long coding and document tasks on K3; large-repo bug fixes fallback to Fable 5; terminal-heavy agents keep GPT-5.6 Sol.
6. Calendar July 27: Track Hugging Face MXFP4/NVFP4 quant builds and vLLM/SGLang day-0 support if you need self-hosting.

July 27 open-weight release: K3 becomes the largest downloadable open model, the first above 2T parameters, and a new fine-tuning foundation — with MXFP4 training enabling efficient inference on modern hardware. Day-0 support in transformers, vLLM, and SGLang is expected.

Citable facts (EEAT):

• Scale: 2.8T total parameters, 16/896 active experts — ~75% above DeepSeek V4 Pro (Moonshot blog, 2026-07-16).
• KDA efficiency: 75% KV cache reduction, up to 6.3× decode speedup at 1M tokens (Moonshot technical docs).
• Intelligence index: 57.1 on Artificial Analysis v4.1 — top-tier among open models (Artificial Analysis, July 2026).

FAQ:

• Is Kimi K3 free? Yes on kimi.com with a free account; API usage is pay-per-token ($3/$15 per 1M).
• Can I run it locally? Not until July 27 weights; then you need 64+ H100-class GPUs for production inference.
• Kimi K3 vs DeepSeek V4 Pro? K3 has nearly 2× parameters, 1M vs 128K context, stronger coding scores — DeepSeek is much cheaper on output ($3.48/M).
• Is 1M context actually useful? Yes for whole-codebase calls, end-to-end legal/research docs, and long-memory agents — flat pricing makes it usable.
• When are low/high reasoning modes coming? Moonshot says subsequent updates; only max is available now.

Bottom line: Kimi K3 is the most capable open-source AI model shipped to date. It does not win every benchmark — Fable 5 and GPT-5.6 Sol still lead specific tasks — but it is competitive across the board, wins long-horizon coding and document understanding, and ships 1M context at Sonnet-tier pricing. July 27 weights are the story to watch.

Dates to bookmark: Now → try at kimi.com · July 17–20 → WAIC Shanghai · July 27 → Hugging Face weights.

Running Kimi K3 agents on a sleeping laptop breaks long SWE Marathon loops; parallel repos fight for RAM and bandwidth. For production Cursor Agent, Kimi Code, or OpenClaw Gateway 24/7, CALMVPS bare-metal Mac Mini rental is usually the better fit: dedicated Apple Silicon, Metal-native acceleration, ~120-second delivery, offload heavy agent loops to the cloud while you review locally. See pricing.