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

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts (MoE) model from Beijing-based Moonshot AI. It is the world's first open model in the 3T class, surpassing the previous record holder DeepSeek V4 Pro (1.6T) by nearly 75% — roughly 2.7x Xiaomi's open model (1.02T) and more than 7x Alibaba's 397B release.

The model activates 16 of 896 experts per forward pass, carries a 1,048,576-token (1M) context window, and accepts text, image, and video inputs with text output. At launch, reasoning runs at max effort only; low and high modes are promised in later updates.

The last 18 months were rough for Moonshot AI. DeepSeek's rise eroded market share. Kimi K3 is a strategic counterpunch timed for the eve of the 2026 World AI Conference (WAIC) in Shanghai.

• For 9 of the past 12 months, Kimi models held the record for the largest open-source model by parameter count.
• ARR crossed $300M as of June 2026 (from $100M in March and $200M in May).
• The company closed its 6th funding round in 2026 at a $31.5B pre-money valuation.
• API revenue exceeds 70% of total revenue; overseas paid users grew 400%.

This is not a vanity scale play from a struggling lab. Moonshot is a fast-growing, API-first business making a technical sovereignty statement on the global stage.

Kimi Delta Attention (KDA)

Standard full attention scales quadratically with context length. At 1M tokens, KV cache memory alone becomes catastrophic. KDA is a hybrid linear attention mechanism that alternates linear-attention layers and full-attention layers in a 3:1 ratio:

• Three linear layers handle local sequence structure cheaply.
• One full-attention layer preserves global information flow.
• Result: up to 75% less KV cache memory and up to 6.3x faster decoding at 1M-token contexts.
• KDA matches or beats full-attention baselines on short-context, long-context, and reinforcement-learning tasks — no capability tradeoff at launch specs.

Attention Residuals (AttnRes)

Standard residual connections accumulate representations uniformly across depth — early-layer signals get diluted in deeper layers. AttnRes replaces this with selective retrieval across depth, letting the model pull high-value representations from earlier layers instead of inheriting whatever was accumulated. Moonshot reports approximately 25% higher training efficiency at under 2% additional compute cost.

Stable LatentMoE

Activating just 16 of 896 experts (1.8% sparsity) breaks standard routing strategies. Moonshot's supporting stack:

Combined, these advances deliver roughly 2.5x better scaling efficiency versus Kimi K2 — the same compute budget produces a significantly smarter model.

Kimi K3's spec sheet is impressive, but production teams hit these friction points quickly:

• Harness inconsistency: Moonshot benchmarks use Kimi Code for K3, Codex for GPT-5.6 Sol, and Claude Code for Anthropic models. Cross-vendor leaderboard rows are directionally useful, not apples-to-apples — independent reproduction is still ongoing.
• Output pricing gap: At $15/M output tokens, K3 is far above DeepSeek V4 Pro ($3.48/M). High-frequency agent loops that emit long patches can erase cache savings on the input side.
• Only max reasoning at launch: No low/high effort dial means you cannot trade latency for cost on simple subtasks yet.
• Self-hosting cliff: July 27 weights require a 64+ accelerator supernode for production inference — this is an inference-provider model, not a MacBook LLM.
• 1M context temptation: Flat pricing encourages stuffing entire repos into one call. Without disciplined cache keys and prompt templates, token volume — not list price — becomes the bill driver.
• FrontierSWE and HLE gaps: Claude Fable 5 still leads on hardest repo-level bug fixing and deepest reasoning (HLE-Full 53.3 vs K3's 43.5). Teams cannot route everything to K3 without acceptance-test regression.

The tables below reflect Moonshot's self-reported numbers as of July 16, 2026. Treat them as vendor-directional until third-party harnesses converge.

Coding benchmarks

SWE Marathon measures sustained, long-horizon coding — the closest public proxy to "write code for hours." Kimi K3 leads at 42.0, seven points ahead of Claude Fable 5 and three ahead of GPT-5.6 Sol. FrontierSWE is where Fable 5 still dominates (86.6 vs 81.2).

Agent, knowledge, and vision benchmarks

OmniDocBench (mixed text, tables, charts, and images) is where K3's native vision plus 1M context compound — it leads all listed models at 91.1. HLE-Full is Fable 5 territory (53.3 vs 43.5).

Artificial Analysis Intelligence Index

On Artificial Analysis Intelligence Index v4.1, Kimi K3 scores 57.1 — 4th place overall:

1. Claude Fable 5 w/ fallback: 59.9
2. GPT-5.6 Sol (max): 58.9
3. GPT-5.6 Sol (xhigh): 57.6
4. Kimi K3: 57.1

The gap between #1 and #4 is just 2.8 points. For an open-weight-bound model competing with Anthropic and OpenAI flagships, that is a remarkable placement — even before July 27 weights land.

Kimi K3 matches Claude Sonnet 5 standard pricing ($3/$15) but delivers 5x the context window. The cache story matters: Moonshot reports 90%+ cache hit rates in Kimi Code workflows via Mooncake split-inference architecture. Effective average input cost can drop to roughly $0.55/M tokens in real usage — OpenRouter's seven-day weighted average corroborates this empirically.

Domestic API pricing (China): ¥20/M input, ¥100/M output, ¥2/M cache-hit input. Consumer access at kimi.com remains free-tier friendly; prepaid bundles start at ¥199 (promotional pricing through August 11, 2026).

Versus Claude Opus 4.8, K3 beats it on several benchmarks at 60% of input cost and 40% of output cost — while offering a context window Opus cannot match at this price.

Option 1 — kimi.com (simplest): Visit kimi.com, sign up with Google. K3 runs at max reasoning effort immediately. No credit card required.

Option 2 — Moonshot API (developers): OpenAI-compatible endpoint at https://api.moonshot.ai/v1. Get keys at platform.kimi.ai.

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)

Option 3 — OpenRouter: Model ID moonshotai/kimi-k3 at Moonshot's official $3/$15 pricing with no markup and full 1M context.

Option 4 — Wait for open weights (July 27): Full model weights release on Hugging Face July 27, 2026. Production deployment requires a supernode with 64+ accelerators — an inference-provider scale, not a laptop deployment.

Moonshot committed to releasing full model weights on July 27, 2026 under a Modified MIT license. When weights land, Kimi K3 becomes:

• The largest downloadable open-source model ever released.
• The first open model above 2 trillion parameters.
• A new fine-tuning and research foundation for the open community.

Training used MXFP4 weights and MXFP8 activations — quantization-aware from the start. Expect day-zero MXFP4/NVFP4 quant builds on Hugging Face and first-class support in vLLM and SGLang, following precedent from comparable MoE releases.

Dates to bookmark: Right now — try K3 at kimi.com; July 17–20 — WAIC Shanghai (more announcements expected); July 27 — full weights on Hugging Face.

1. 01
   Create a Moonshot API key: register at platform.kimi.ai, generate a key, and store it in your secrets manager — not in committed source.
2. 02
   Pick an access path: direct Moonshot API for lowest latency, OpenRouter (moonshotai/kimi-k3) if you already route through a gateway, or kimi.com for manual validation before wiring agents.
3. 03
   Configure cache-friendly prompts: reuse system prompts and stable prefix blocks so Mooncake split-inference can hit the $0.30/M cache tier. Target 90%+ cache hits on repetitive agent loops.
4. 04
   Smoke-test with a bounded task: run a 10k-token repo summary before stuffing a full 1M context — confirm quality and latency before scaling to whole-codebase calls.
5. 05
   Wire into your agent harness: point OpenAI-compatible clients at https://api.moonshot.ai/v1 with model="kimi-k3"; log input/output tokens and cache-hit ratio per task in CI.
6. 06
   Build a mixed routing policy: route long-horizon coding and document analysis to K3; reserve FrontierSWE-class bug fixes and HLE-depth reasoning for Claude Fable 5; track per-task cost on the pricing page assumptions before committing volume.

Kimi K3 is the most capable open-weight-bound AI model released to date. It does not win every benchmark — Claude Fable 5 and GPT-5.6 Sol still lead on specific coding and reasoning tasks — but it is competitive across the board, outperforms them on long-horizon coding and document understanding, and ships a 1M-token context window that closed-source competitors cannot match at this price.

The July 27 weight release is the story to watch. If Moonshot delivers Modified MIT weights with vLLM and SGLang day-zero support, K3 could reshape the open-source landscape for the rest of 2026.

Teams adopting K3 for agent workflows against large codebases still need a stable local dev and CI plane. Running agents on shared minute pools, home Macs, or oversubscribed VPS hosts introduces bandwidth jitter, neighbor CPU contention, and dropped long-lived SSH sessions — friction that can erase the token savings from cheaper models. For production agent hosts and auditable build environments, NUKCLOUD multi-region bare-metal Mac / cloud Mac nodes offer dedicated Apple Silicon compute with clear tenant boundaries. Compare specs on the pricing page and provision a trial via order.

• Is Kimi K3 available for free?
  Yes on kimi.com with a free account. API usage requires a paid key at $3/$15 per million tokens (input/output).
• Can I run Kimi K3 locally?
  Not until July 27, 2026, when full weights release on Hugging Face. Production inference needs 64+ accelerators — not a consumer Mac deployment. For local MoE experimentation today, see our DeepSeek V4 Mac cloud rental runbook.
• How does Kimi K3 compare to DeepSeek V4 Pro?
  K3 has nearly double the parameters (2.8T vs 1.6T), 8x the context (1M vs 128K), and stronger benchmark scores on most coding tests — but DeepSeek output costs $3.48/M versus K3's $15/M.
• Is the 1M token context window actually useful?
  Yes for whole-codebase analysis, end-to-end document review, and multi-session agents. Flat pricing plus 90%+ cache hits in Kimi Code make using the full window practical.
• When are low and high reasoning effort modes coming?
  Moonshot says low and high modes arrive in subsequent updates. Only max effort is available at launch.
• How does Kimi K3 compare to Claude Fable 5?
  Fable 5 leads FrontierSWE and HLE-Full and scores 59.9 vs K3's 57.1 on Artificial Analysis. K3 leads SWE Marathon, Program Bench, BrowseComp, Automation Bench, and OmniDocBench — with Sonnet-level pricing and 5x the context.
• Where should I run agent CI against large codebases?
  For stable local dev and CI when agents hammer large repos over long Kimi K3 sessions, NUKCLOUD bare-metal Mac nodes provide dedicated compute without the jitter and connection drops common on shared pools — pair API savings with a reliable build plane.

Data as of 2026-07-17. Sources: Moonshot AI official blog, Kimi API Platform docs, Artificial Analysis Intelligence Index v4.1, OpenRouter pricing. Benchmarks are self-reported by Moonshot AI as of July 16, 2026.