What Is Kimi K3? Core Specs of the 2.8-Trillion-Parameter Open-Source Model
On July 16, 2026, Moonshot AI released Kimi K3 — a 2.8-trillion (2.8T) sparse Mixture-of-Experts (MoE) model that surpasses the previous record holder DeepSeek V4 Pro (1.6T) by nearly 75%, is 2.7x the size of Xiaomi's open model (1.02T), and more than 7x Alibaba's (397B). The quiet rollout contrasts sharply with the scale: one technical blog post, one pricing page, and a model ID you can call immediately — kimi-k3.
K3 uses 896 experts with 16 activated per inference (1.8% sparsity); paired with a 1-million-token ultra-long context (roughly the full text of five novels in one pass) and native vision understanding, it is built for complex coding, long-document reasoning, and knowledge work. In one line: an open-source, natively multimodal coding AI with ultra-long memory, priced 40% below Claude Opus 4.8, with full weights open-sourcing on July 27.
Scale shock: After DeepSeek's rise eroded Moonshot's market share, K3 is a strategic counter — the Kimi family held the open-source size ceiling for 9 of the past 12 months.
Launch timing: Dropped the night before the 2026 World AI Conference (WAIC), July 17–20, with more announcements expected.
Commercial surge: ARR passed $300M by June 2026; a sixth funding round this year at a $31.5B pre-money valuation; API revenue exceeds 70% of total, with overseas paid users up 400%.
Long-context pain point: Full attention at 1M tokens causes KV cache memory to explode — K3's KDA mechanism exists specifically to solve this.
Coding Agent demand: Long-horizon tasks like SWE Marathon require models to stay coherent over multi-hour sessions — the 1M window plus 90%+ cache hit rate is a core selling point.
| Parameter | Value |
|---|---|
| Total parameters | 2.8 trillion (world's largest open-source model) |
| Architecture | KDA + AttnRes + Stable LatentMoE |
| Active experts | 16 / 896 (sparse MoE, 1.8%) |
| Context window | 1,048,576 tokens (1M) |
| Input modalities | Text, image, video |
| Reasoning mode | Always-on max effort (low/high coming in future update) |
| Open weights | July 27, 2026 on Hugging Face |
Kimi K3 Architecture and Benchmarks: KDA, AttnRes, and Stable LatentMoE
Kimi Delta Attention (KDA) is a hybrid linear attention mechanism: it alternates linear and full attention layers in a 3:1 ratio — three linear layers handle local structure (cheap compute), one full attention layer preserves global information flow. Result: KV cache memory drops by up to 75%, decode speed at 1M tokens improves by up to 6.3x, and it beats pure full-attention baselines on both short and long context and RL scaling scenarios.
Attention Residuals (AttnRes) rework depth-dimension residual connections: standard residuals dilute critical representations from early layers; AttnRes adds selective retrieval so the model can pull high-value representations from earlier layers across depth, delivering roughly 25% training efficiency gain with under 2% extra compute.
Stable LatentMoE enables stable training at extreme sparsity (896 experts / 16 active). Supporting techniques:
| Technique | Role |
|---|---|
| Quantile Balancing | Derives expert allocation directly from router score quantiles, eliminating heuristic hyperparameters |
| Per-Head Muon | Independent optimization per attention head for more adaptive large-scale training |
| Sigmoid Tanh Unit (SiTU) | Improved activation function control |
| Gated MLA | Enhanced attention selectivity |
Combined, these innovations raise K3's overall scaling efficiency by roughly 2.5x versus Kimi K2. Moonshot's self-reported benchmark data below (each model uses its own reasoning harness; independent third-party replication is still underway):
| 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 (vision) | 81.6 | 81.2 | 83.0 | 78.9 | — |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 | — |
SWE Marathon tests sustained long-horizon coding — K3 leads at 42.0 by a wide margin. On Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 (fourth overall), trailing Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9) by just 2.8 points.
Kimi K3 Pricing: How Much Does It Cost vs Claude, GPT, and DeepSeek?
K3 standard pricing matches Claude Sonnet 5 ($3/$15 per 1M tokens) but offers a 5x larger context window (1M vs 200K). Cache-hit input drops to $0.30/M (one-tenth of standard); Moonshot reports over 90% cache hit rates in coding workflows, making effective input cost extremely low. OpenRouter's 7-day weighted average confirms effective input cost around $0.55/M.
| Model | Input ($/M) | Output ($/M) | Cache-hit input | Context |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 | 1M |
| Claude Sonnet 5 | $3.00 (promo $2) | $15.00 (promo $10) | — | 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 |
| Kimi K2.6 | $0.95 | $4.00 | $0.16 | 256K |
Versus Claude Opus 4.8, K3 beats it on multiple benchmarks while input costs 60% and output costs 40% as much. China API pricing: input ¥20/M, output ¥100/M, cache-hit ¥2/M. Consumer tier: free account on Kimi.com; prepaid plans from ¥199 (promo through August 11).
How to Use Kimi K3 Now: Four Access Methods and Six-Step Runbook
Kimi web/app (simplest): Visit kimi.com, register (Google sign-in supported). K3 runs at max reasoning effort by default — no credit card required.
Get an API key: Register a developer account at platform.kimi.ai, create an API key, and add credits.
Configure OpenAI-compatible client: Set base_url="https://api.moonshot.ai/v1", model ID kimi-k3.
OpenRouter: Model ID moonshotai/kimi-k3, Moonshot official pricing with no markup, full 1M context.
Enable prompt caching: Set conversation routing cache keys in coding workflows to leverage 90%+ cache hit rates and cut effective input cost to ~$0.55/M.
Mark July 27: Full model weights release on Hugging Face; MXFP4/NVFP4 quantized versions and Day-0 vLLM/SGLang support expected. Production deployment needs a 64+ accelerator supernode.
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 code for me..."}]
)
Open-source timeline: WAIC July 17–20 may bring more announcements, then July 27 full K3 weights open-source — the first downloadable open weights above 2 trillion parameters.
Which Model Should You Pick? Use Case Matrix and Citable Data
| Scenario | Recommended model | Why |
|---|---|---|
| Sustained long coding tasks | Kimi K3 | SWE Marathon #1, longest 1M context |
| Complex repo-level bug fixes | Claude Fable 5 | FrontierSWE 86.6 leads by a wide margin |
| Terminal/toolchain-heavy Agents | GPT-5.6 Sol | Terminal Bench 2.1 and Coding Agent Index lead |
| Ultra-long docs / multimodal understanding | Kimi K3 | OmniDocBench 91.1 #1, native vision + 1M context |
| Cost-sensitive workloads | DeepSeek V4 Pro | Output only $3.48/M, far below K3 |
| Self-hosted open weights (post 7/27) | Kimi K3 | Strongest open weights, MXFP4 quantization friendly |
Parameter scale: 2.8T — nearly 75% above DeepSeek V4 Pro (1.6T), largest downloadable open model (after 7/27).
KDA efficiency: KV cache -75%, 1M-token decode +6.3x, training scaling efficiency +2.5x (vs K2).
Intelligence Index: Artificial Analysis v4.1 score 57.1 — highest among open models, 2.8 points behind closed-source flagships.
Caveat: Benchmarks above are Moonshot self-reported. K3 uses Kimi Code, GPT uses Codex, Claude uses Claude Code — each with its own harness. Independent third-party replication is still underway; treat as directional guidance.
Kimi K3 is not parameter vanity — KDA, AttnRes, and Stable LatentMoE are real engineering innovations that match or beat some closed-source flagships on long coding tasks and document understanding. API-only setups work for quick starts, but local Macs often hit memory pressure under long Kimi Code Agent loops, and VMs sacrifice Metal/Xcode fidelity. For 7x24 coding Agent uptime, parallel dev servers, and always-on inference, MESHLAUNCH cloud Mac Mini bare-metal rental is usually the better production fit: dedicated Apple Silicon with flexible daily/weekly/monthly terms, ideal for iOS CI/CD and AI Agent automation.
Sources: Moonshot AI official blog · Kimi API Platform docs · Artificial Analysis · OpenRouter pricing · VentureBeat · SCMP
Yes. Register a free account at kimi.com to use K3 at max reasoning effort. API calls are paid ($3/$15 per 1M tokens). Budget planning: see our pricing page.
Full weights release on Hugging Face July 27, 2026. Production deployment requires a 64+ accelerator supernode; consumer-grade local deployment is not practical. Model trained with MXFP4 weights + MXFP8 activations — quantization friendly.
K3 has nearly 2x parameters, 8x larger context, and stronger coding and document benchmarks. DeepSeek V4 Pro output costs only $3.48/M — better for cost-sensitive workloads. See the help center for related guidance.
Very useful for analyzing entire codebases in one pass, processing long legal or research documents, and multi-turn Agent long-memory scenarios. Flat pricing with no length surcharge, plus 90%+ cache hit rates, keeps real costs manageable.
Moonshot AI says low and high effort modes will roll out in a future update. Only max mode is available today.