KIMI K3
2.8T_OPEN_
SOURCE_REVIEW.

Kimi K3 2.8T open-source LLM architecture and benchmark comparison

Lead: Developers evaluating closed-source APIs keep hitting the same three walls: Is the context window long enough? Can it actually code? When can I self-host the weights? On the night of July 16, 2026, Moonshot AI quietly flipped a switch — a banner at the top of the Kimi API docs read "Kimi K3 is live!" This review covers every material point from our research: 2.8T parameters, the KDA / AttnRes / Stable LatentMoE architecture stack, a full benchmark table against Claude Fable 5 and GPT-5.6 Sol, pricing matrices, five integration paths, the July 27 open-weight release, a scenario selection matrix, and FAQ.

30-Second Read · TL;DR

Launch2026-07-16 quiet midnight release · no press event
Scale2.8 trillion parameters · world's largest open model · ~75% above DeepSeek V4 Pro (1.6T)
Core capability1M token context · native vision · MoE with 16 of 896 experts active
PricingInput $3/M · output $15/M · cache hit $0.30/M
Open weightsFull weights on Hugging Face 2026-07-27

1. Pain Points: Why K3 Deserves Your Attention

  1. Moonshot needed a technical sovereignty statement after DeepSeek's rise. Over the past 18 months, DeepSeek eroded Moonshot's market position significantly. K3 reclaims the "largest open-source model" title at 2.8T parameters — timed for the eve of the 2026 World AI Conference (WAIC) in Shanghai.
  2. Long context is no longer a paper spec. Standard full attention makes KV cache memory catastrophic at 1M tokens. Kimi Delta Attention (KDA) cuts KV cache by up to 75% and boosts million-token decode speed by up to 6.3× — making 1M context practical at a flat $3/M input price.
  3. "Can code" and "can code for hours" are different problems. SWE Marathon tests sustained long-horizon coding work. K3 scores 42.0, well ahead of Claude Fable 5 (35.0) and GPT-5.6 Sol (39.0). If your workflow involves multi-hour coding sessions, this benchmark is closer to reality than a single SWE-bench pass.

2. What Is Kimi K3?

On the night of July 16, 2026, Moonshot AI quietly launched Kimi K3. No press conference. No countdown timer. Just a technical blog post, a pricing page, and a model ID you could call immediately: kimi-k3.

Kimi K3 is the world's largest open-source AI model by parameter count — 2.8 trillion (2.8T) parameters, surpassing the previous record holder DeepSeek V4 Pro (1.6T) by nearly 75%. It is 2.7× the size of Xiaomi's open model (1.02T) and more than 7× Alibaba's open offering (397B).

It uses a sparse mixture-of-experts (MoE) architecture, activating 16 of 896 experts per forward pass. Combined with a 1-million-token context window (roughly five full novels of text in a single pass) and native vision understanding, it is built for complex coding tasks, long-document reasoning, and knowledge work.

One-line summary: Kimi K3 is an open-source, vision-native, long-memory coding AI that costs 40% less than Claude Opus 4.8 on output, with full weights releasing July 27.

SpecDetail
Total parameters2.8 trillion
ArchitectureKDA + Attention Residuals + Stable LatentMoE
Active experts16 of 896 (1.8% sparsity)
Context window1,048,576 tokens (1M)
Input modalitiesText, image, video
ReasoningAlways-on, max effort at launch
API model IDkimi-k3
Open weightsJuly 27, 2026

3. Background: Why This Release Matters

  • For 9 of the past 12 months, Kimi models held the record for the largest open-source model by parameter count
  • Kimi K3 launched on the eve of the 2026 World AI Conference (WAIC) in Shanghai
  • As of June 2026, Moonshot AI's ARR crossed $300M, with a 6th funding round this year at a $31.5B pre-money valuation
  • API revenue makes up over 70% of total revenue; overseas paid users grew 400%

This is not a vanity project from a company burning cash on scale. Moonshot is a fast-growing business making a serious technical statement. For context on the closed-source frontier, see our coverage of Claude Sonnet 5 / GPT-5.6 leak intelligence and GPT-5.6 Sol Ultra mathematical proof analysis.

4. Core Architecture: Three Genuine Innovations

4.1 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 layers and full-attention layers in a 3:1 ratio — three cheap linear layers handle local structure, one full-attention layer preserves global information flow
  • KV cache memory reduced by up to 75%
  • Decode speed at 1M-token contexts improved by up to 6.3×
  • Matches or beats full-attention baselines on short-context, long-context, and reinforcement learning tasks — no capability tradeoff
Simple analogy: full attention is like trying to remember every detail of every conversation simultaneously. KDA is more like an efficient assistant — fast indexing most of the time, precise recall when it matters.

4.2 Attention Residuals (AttnRes) — Fixing Information Loss Across Depth

Standard residual connections accumulate representations uniformly across depth — early-layer signals get diluted as you go deeper. AttnRes introduces selective retrieval across depth: the model can reach back and pull high-value representations from earlier layers rather than inheriting whatever was accumulated. Moonshot reports approximately 25% higher training efficiency at under 2% additional compute cost.

4.3 Stable LatentMoE — Making Extreme Sparsity Actually Work

Kimi K3 has 896 experts, activating 16 per forward pass — a sparsity of 1.8%. At this level, standard routing strategies break down. Supporting techniques:

TechniqueRole
Quantile BalancingDerives expert allocation from router-score quantiles, eliminating fragile heuristic hyperparameters
Per-Head MuonOptimizes each attention head independently for more adaptive learning at scale
Sigmoid Tanh Unit (SiTU)Improved activation function control
Gated MLAEnhanced attention selectivity

Combined, these advances deliver approximately 2.5× better scaling efficiency compared to Kimi K2 — the same compute budget produces a significantly smarter model.

5. Benchmarks: Where It Wins, Where It Doesn't

BenchmarkKimi K3Claude Fable 5GPT-5.6 SolClaude Opus 4.8GLM-5.2
DeepSWE67.570.073.059.046.2
Program Bench77.876.877.671.963.7
Terminal Bench 2.188.384.688.884.682.7
FrontierSWE81.286.671.366.767.3
SWE Marathon42.035.039.040.013.0
BrowseComp91.288.090.484.3
Automation Bench30.829.129.727.212.9
GPQA-Diamond93.592.694.191.091.2
MMMU-Pro (vision)81.681.283.078.9
OmniDocBench (document)91.189.885.887.9

Key takeaways:

  • SWE Marathon: K3 leads at 42.0 — the benchmark closest to "actually writing code for hours"
  • Program Bench: K3 edges first place at 77.8
  • FrontierSWE: Fable 5 leads (86.6); K3 (81.2) still crushes GPT-5.6 Sol (71.3)
  • OmniDocBench: K3 first at 91.1 — vision + long context working together
  • Overall intelligence: On Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 (4th overall), behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9) — a gap of just 2.8 points to #1

Caveat: These are Moonshot's self-reported benchmarks. Different harnesses were used (Kimi Code for K3, Codex for GPT, Claude Code for Claude). Independent third-party reproductions are still ongoing. Treat these as directionally useful, not definitive.

6. Pricing: Cheaper Than Claude Opus, Parity With Sonnet

ModelInput ($/M)Output ($/M)Cache-hit inputContext
Kimi K3$3.00$15.00$0.301M
Claude Sonnet 5$3.00 (promo $2)$15.00 (promo $10)200K
Claude Opus 4.8$5.00$25.00200K
GPT-5.5$5.00$30.00400K
DeepSeek V4 Pro$1.74$3.48$0.145128K
Kimi K2.6$0.95$4.00$0.16256K
  • K3 matches Claude Sonnet 5 standard pricing ($3/$15) but offers the context window
  • Cache-hit input drops to $0.30/M; Moonshot reports 90%+ cache hit rates in coding workflows via Mooncake split-inference architecture — effective average input cost can fall to roughly $0.55/M in real usage
  • Compared to Claude Opus 4.8, K3 beats it on several benchmarks at 60% of input cost and 40% of output cost
  • Free tier available at kimi.com; API requires pay-per-token billing

7. Getting Started: Five Ways to Use K3 Today

  1. Kimi web / app (simplest): Visit kimi.com, sign up with Google. K3 runs at max reasoning effort by default. No credit card required.
  2. Official API (developers): Get an API key at platform.kimi.ai. Fully OpenAI-compatible:
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)
  1. OpenRouter: Model ID moonshotai/kimi-k3 — official Moonshot pricing, no markup, full 1M context.
  2. Cursor / IDE multi-routing: Set Kimi K3 as your long-context coding primary; fall back to Claude Fable 5 for complex repo-level bug fixes.
  3. Wait for July 27 open weights: Full model weights on Hugging Face (requires a 64+ accelerator supernode for production deployment — not a laptop LLM).

8. Comparison Matrix: Which Model Should You Use?

Use caseBest pickWhy
Long, sustained coding sessionsKimi K3Leads SWE Marathon; 1M context prevents mid-task context loss
Complex repo-level bug fixesClaude Fable 5FrontierSWE leads by a wide margin
Terminal / tool-heavy agent workflowsGPT-5.6 SolLeads Terminal Bench and Coding Agent Index
Multimodal document analysisKimi K3Best OmniDocBench score; native vision + 1M context
Cost-sensitive production useDeepSeek V4 ProOutput at $3.48/M, far cheaper
Open-source self-hosting (post 7/27)Kimi K3Most capable open weights available
Deepest reasoning researchClaude Fable 5Leads HLE-Full by a wide margin

9. Open-Source Promise: What Happens July 27?

Moonshot AI has committed publicly: full model weights release on July 27. When they drop, Kimi K3 will become:

  • 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

The model was trained with MXFP4 weights and MXFP8 activations — quantization-aware from the start. Expect MXFP4/NVFP4 quantized versions on Hugging Face, with day-0 support in transformers, vLLM, and SGLang.

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

10. Frequently Asked Questions

Q: Is Kimi K3 available for free?
A: Yes — on kimi.com with a free account. API usage requires a paid API key at $3/$15 per million tokens.

Q: Can I run Kimi K3 locally?
A: Not yet — weights release July 27, 2026. Production deployment needs 64+ accelerators. Not practical on consumer hardware at 2.8T parameters.

Q: How does Kimi K3 compare to DeepSeek V4 Pro?
A: K3 has nearly double the parameters (2.8T vs 1.6T), 1M vs 128K context, and stronger benchmarks — but DeepSeek output is $3.48/M vs K3's $15/M.

Q: Is the 1M token context window actually useful?
A: Yes — for full-codebase analysis, long legal/research documents, and multi-session agent memory. Flat pricing means no length surcharge.

Q: When are low/high reasoning effort modes coming?
A: Moonshot says low and high modes are coming in subsequent updates. Currently only max is available.

11. Deep Insight: A Turning Point for Open-Source AI

Kimi K3 is not a parameter-stacking vanity project. The architecture work — KDA, AttnRes, Stable LatentMoE — addresses real engineering problems at scale. On long-horizon coding and document understanding, it matches or beats parts of the closed-source frontier at reasonable pricing, with a credible open-weight commitment.

More broadly, this signals a shift in China's open-source AI ecosystem: no longer competing purely on price, but challenging the intelligence frontier directly. When July 27 weights land, the open-source community gets its first 2T+ downloadable foundation model — potentially the most significant open release of 2026 for teams building agents, fine-tuning vertical models, or deploying inside corporate networks without closed API dependency.

For Mac developers, routing K3 through Cursor for long-context coding and running Kimi Code on SWE Marathon-class tasks is the pragmatic path today. After weights drop, tracking MLX and llama.cpp community quantization progress will matter — the same way infrastructure competition shaped the DeepSeek ecosystem.

12. Closing: API on Any Device, Agent Testing Still Belongs on Mac

Signing up at kimi.com, copying an API key, and switching routes on OpenRouter — Windows or Linux works fine for that. But if you want to run Cursor + Kimi K3 long-context coding, Xcode-side agent integration, or MLX-based quantization validation on the same machine, Apple Silicon unified memory + Metal remains the lowest-friction path.

The practical split: keep your daily driver on API calls, and offload Kimi Code stress tests, multi-repo SWE Marathon regressions, and 1M-token document batch jobs to a MACGPU remote Mac mini M4 node — on-demand, SSH-secured. Before K3 weights land July 27, isolate agent workloads so your primary machine stays stable for everyday development.