When Moonshot AI dropped Kimi K3 on July 16, 2026, the AI world collectively paused. Here was a Chinese lab doing something no one had done before — releasing a 2.8-trillion-parameter open-weight model that could spar with the best closed models from Anthropic and OpenAI. I have spent the past week digging through benchmarks, pricing tables, developer forums, and official documentation to bring you a review that goes beyond the headlines. Whether you are a developer weighing API costs, a researcher curious about multimodal capabilities, or a business owner evaluating AI tools, this guide breaks down everything you need to know about Kimi K3.
At Scoptechs, we have been tracking the AI revolution closely — from our coverage of AI tools to deep dives on emerging technologies. Kimi K3 represents a pivotal moment, and I want to help you understand exactly where it fits in the landscape.
What Is Kimi K3? A Quick Overview
Kimi K3 is the flagship large language model from Moonshot AI, the Beijing-based lab backed by Alibaba, Tencent, and Meituan. It arrived as a Mixture-of-Experts (MoE) multimodal model with a staggering 2.8 trillion total parameters, activating only 16 of its 896 experts per token to keep inference efficient. The model ships with a 1,048,576-token context window — that is one million tokens of context, flat-priced with no long-context surcharge — and it processes text, images, and video natively.
What sets K3 apart from its K2 predecessors is not just scale. Moonshot rebuilt the attention mechanism from the ground up, introduced always-on reasoning with a tunable effort dial, and committed to releasing the full model weights under a Modified MIT license. As of this writing, those weights are promised by July 27, 2026, which would make K3 the largest openly available AI model in history.
Key Facts at a Glance
- Developer: Moonshot AI (Beijing, China)
- Release Date: July 16, 2026
- Total Parameters: 2.8 trillion (MoE architecture)
- Active Experts: 16 out of 896 per token
- Context Window: 1,048,576 tokens (1M)
- Modalities: Text, images, video (native, no bolt-on OCR)
- Reasoning: Always-on with tunable
reasoning_effort(low/medium/high) - License: Modified MIT (weights pending release)
- API Pricing: $0.30/MTok cached input · $3.00/MTok uncached input · $15.00/MTok output
Under the Hood: Architecture and Technical Innovation
If you want to understand why Kimi K3 punches above its weight, you need to look at three architectural decisions Moonshot made. These are not marketing buzzwords — they directly affect how the model performs on real tasks.
Pictures inside kimi k3
Kimi Delta Attention (KDA)
Traditional attention mechanisms scale quadratically with sequence length, which means longer contexts become computationally expensive. K3 introduces Kimi Delta Attention, a hybrid linear attention mechanism that allows information to flow more smoothly across very long sequences. Moonshot claims this delivers roughly a 2.5x scaling-efficiency improvement over K2. While that is a vendor-reported figure, the practical consequence is verifiable: K3 can handle its full million-token window without the performance cliff that older architectures hit around 128K or 256K tokens.
Attention Residuals (AttnRes)
Deep models often suffer from information loss as data passes through dozens of layers. Attention Residuals address this by preserving and re-injecting information from earlier layers, which helps the model maintain coherence on long-horizon tasks like multi-file code refactoring or extended document analysis.
Stable LatentMoE Framework
The Mixture-of-Experts design activates only 16 of 896 experts per token. This extreme sparsity is what allows a 2.8T model to serve responses at a price point below the closed frontier leaders. Think of it as having a massive team of specialists but only calling in the relevant ones for each specific task.
Benchmark Performance: How Does K3 Actually Stack Up?
Benchmarks tell part of the story, but they require context. Moonshot ran K3 through an extensive evaluation suite, and third-party evaluators like Artificial Analysis have independently verified several scores. Here is what the numbers say.
Coding Benchmarks (K3's Strongest Category)
Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
SWE-bench Verified | 76.8% | 79.1% | 78.5% |
Terminal-Bench 2.1 | 88.3% | 84.6% | 88.8% |
FrontierSWE | 81.2 | 86.1 | 83.0 |
Program Bench | 77.8 | 77.3 | 77.6 |
SWE Marathon | 42.0 | 35.1 | 40.0 |
DeepSWE | 67.5 | 70.2 | 73.0 |
Kimi Code Bench 2.0 | 72.9 | — | — |
K3 takes first place on Program Bench, SWE Marathon, and Terminal-Bench 2.1 (where it edges out Fable 5 by nearly four points). It trails on FrontierSWE and DeepSWE, where Fable 5 and GPT-5.6 Sol pull ahead. The SWE Marathon lead is particularly noteworthy — this benchmark measures sustained software engineering over long sessions, which is exactly the agentic coding scenario K3 was built for.
Knowledge and Reasoning Benchmarks
Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
GPQA Diamond | 93.5% | 92.6% | 94.1% |
HLE-Full (no tools) | 43.5% | 53.3% | 48.7% |
HLE-Full (with tools) | 56.0% | 61.2% | 58.4% |
MMMU-Pro | 81.6 | 83.9 | 82.1 |
BrowseComp | 91.2 | 90.4 | 91.0 |
OmniDocBench | 91.1 | 89.8 | 90.3 |
K3 leads on BrowseComp (the best browsing-agent score ever published) and OmniDocBench (document understanding), and it sits comfortably between Fable 5 and GPT-5.6 Sol on GPQA Diamond. The gap on HLE-Full — Humanity's Last Exam — is where K3 shows its clearest weakness in raw reasoning depth.
A Critical Caveat About Harness Differences
One thing every reviewer should mention: these scores were collected under different agent harnesses. K3 uses KimiCode, Fable 5 uses Claude Code, and GPT-5.6 Sol uses Codex. The harness is part of the system being tested, not just the model. Moonshot documents this openly, which is a refreshing change from typical model launches. Independent verification from Artificial Analysis places K3 at 85% on Terminal-Bench (vs. the 88.3% vendor-reported), which is still strong but shows the harness effect is real.
Pricing: How Much Does Kimi K3 Actually Cost?
Pricing is where K3 becomes genuinely disruptive. Let me break down the three access tiers because the headline number ($3/$15 per million tokens) only tells part of the story.
Tier 1: Free Web Chat
You can use K3 for free at kimi.com with no account required for your first message. This is the zero-risk way to test the model.
Tier 2: Consumer Subscription Plans
Kimi offers subscription tiers ranging from $19 to $199 per month, which include credits for K3 and K3 Swarm access. If you are a power user who wants the agentic features without managing API billing, this is the sweet spot.
Tier 3: API Pricing (Pay-Per-Token)
Pricing Component | Cost |
|---|---|
Cached Input Tokens | $0.30 per million |
Uncached Input Tokens | $3.00 per million |
Output Tokens | $15.00 per million |
Blended Rate (7:2:1 ratio) | $2.31 per million |
The cache pricing is the secret weapon. When your prompt prefix hits cache, input costs drop by 90%. For applications that reuse system prompts or context windows — which is most production use cases — this makes K3 dramatically cheaper than the headline rate suggests.
Price Comparison vs. Competitors
Model | Input (per MTok) | Output (per MTok) |
|---|---|---|
Kimi K3 | $3.00 | $15.00 |
GPT-5.6 Sol | $5.00 | $30.00 |
Claude Fable 5 | $10.00 | $50.00 |
At list price, K3's input is 40% cheaper than Sol and 70% cheaper than Fable 5. Output pricing shows even bigger savings — 50% cheaper than Sol and 70% cheaper than Fable 5. However, there is a catch: K3 is verbose. Artificial Analysis measured it generating 130M tokens in evaluation versus a 63M average, and those extra output tokens are billed at the full output rate. Always factor token volume into your cost calculations.
Real-World Use Cases: Where K3 Shines and Where It Stumbles
Where K3 Excels
Long-Horizon Coding: If you need a model that can sustain a multi-hour coding session, understand a large codebase, and coordinate terminal tools with minimal supervision, K3 is arguably the best open model available. Its SWE Marathon lead (42.0 vs. 40.0 for Sol) directly reflects this capability.
Frontend Development: K3 ranked first on Arena.ai's Frontend Code Arena with 1,679 Elo points, ahead of Claude Fable 5 and GPT-5.6 Sol. It leverages screenshots and visual feedback to debug UI code, which is a game-changer for frontend engineers.
Document Processing: With native multimodal understanding (not bolt-on OCR), K3 can parse PDFs, Word files, Excel spreadsheets, and even write formulas or generate PPT outlines. Its OmniDocBench score of 91.1 leads the field.
3D Web Design and Creative Coding: Early testers report impressive results using K3 for 3D web design, game development, and CAD workflows. The model can inspect rendered results rather than relying solely on source code, making it relevant for visual engineering tasks.
Where K3 Falls Short
Raw Reasoning Depth: On Humanity's Last Exam, K3 trails Fable 5 by nearly 10 points. For tasks requiring deep, multi-step logical reasoning, the closed frontier models still hold an edge.
Hallucination Rate: K3 attempts more answers but also gets more wrong. On AA-Omniscience, accuracy rose from 33% to 46% compared to K2, but hallucination rate also climbed from 39% to 51%. This is the trade-off of an aggressive, confident model — verify its outputs on factual tasks.
Speed: Artificial Analysis flags K3 as "notably slow" compared to peers. The always-on reasoning means every request runs a full thinking pass before answering, which is great for hard problems but wasteful for simple queries.
Infographic: Kimi K3 at a Glance
╔══════════════════════════════════════════════════════════════════╗
Mind Map: How Kimi K3 Fits Together
┌─────────────────────┐
Mind Map: How Kimi K3 Fits
Strengths and Weaknesses Summary
✅ Strengths