Kimi K3 (Moonshot AI)

Learning Objectives

• Understand how Kimi K3's sparse architecture reaches 2.8 trillion parameters while activating only a fraction per token
• Identify the benchmarks where K3 leads Claude Opus 4.8 and where it still trails Claude Fable 5 and GPT-5.6 Sol
• Evaluate when an open-weights frontier model is the right choice over a proprietary API

What Is Kimi K3?

Kimi K3 is the flagship foundation model from Moonshot AI (月之暗面), the Beijing-based lab founded in 2023 by Yang Zhilin. Moonshot launched K3 on July 16, 2026 across Kimi.com, the Kimi App, Kimi Work, Kimi Code, and the Kimi API Platform, and committed to releasing the full model weights by July 27, 2026.

At 2.8 trillion total parameters, K3 is the largest open-weights model any Chinese lab has produced — and its significance is less about raw size than about position. K3 is the first open-weights model to land credibly inside the frontier conversation rather than a generation behind it: it outperforms Claude Opus 4.8 on most of the coding and agentic suite Moonshot published, while Moonshot itself concedes the model still trails Claude Fable 5 and GPT-5.6 Sol overall.

Architecture & Specifications

K3 is a mixture-of-experts (MoE) model built on three architectural components Moonshot introduced with this generation:

• Stable LatentMoE — the routing framework that manages extreme sparsity, activating just 16 of 896 experts per token. That ratio is far sparser than the K2 family, which activated roughly 32 billion of 1 trillion parameters.
• Kimi Delta Attention (KDA) — an attention mechanism Moonshot describes as "an efficient foundation for scaling attention," and the basis for the 1 million token context window.
• Attention Residuals (AttnRes) — a depth-wise mechanism that "selectively retrieves representations across depth rather than accumulating them uniformly."

K3 also ships with native vision rather than a bolted-on adapter, which shows up in its multimodal scores below.

Benchmark Performance

Moonshot published K3 against the current frontier cohort — Claude Opus 4.8, GPT-5.6 Sol, and Claude Fable 5. The pattern is consistent: K3 beats Opus 4.8 nearly everywhere, trades with GPT-5.6 Sol, and trails Fable 5 on reasoning.

The headline result is FrontierSWE, where K3 scores 81.2 against Opus 4.8's 66.7 — a wide margin on a hard software-engineering suite. K3 also leads the whole cohort on SWE Marathon, BrowseComp, and OmniDocBench.

The losses are just as informative. On HLE-Full, a reasoning benchmark, K3 scores 43.5 against Fable 5's 53.3 — a real gap that the coding wins do not close. K3 also trails on Toolathlon-Verified and APEX-Agents, both agentic tool-use suites, which cuts against the assumption that a strong coding model is automatically a strong agent.

Why This Release Matters

For roughly two years the open-weights argument carried an implicit concession: open models were cheaper and more private, but a generation behind the frontier. K3 is the strongest case yet that the concession is narrowing. A lab can now download a model that beats the previous quarter's leading proprietary flagship on most coding work, run it in its own jurisdiction, and pay nothing per token.

Two caveats keep this from being a clean story. First, K3 does not beat the current proprietary frontier — Fable 5 and GPT-5.6 Sol are still ahead, and Moonshot says so plainly. Second, "open weights" is not "cheap to run": a 2.8 trillion parameter model is far beyond what a workstation serves, so most teams will consume K3 through a hosted endpoint anyway, which returns the data-residency question that open weights were supposed to answer.

The release also arrived one day before Xi Jinping used the World AI Conference in Shanghai to make open source explicit Chinese national strategy — making K3 the most concrete evidence available for a policy position China's head of state was articulating the next morning.

Pricing & Access

The pricing tells its own story. At $15 per million output tokens, K3 is not undercutting the frontier the way earlier Chinese open models did — it is priced like a frontier model. The aggressive number is the 30 cents per million cached input tokens, which rewards agentic workloads that re-send a large context repeatedly.

Limitations & Considerations

• Preserved thinking history is mandatory: Moonshot warns that K3 "was trained in the preserved thinking history mode" and that "generation quality may become highly unstable" if thinking content is not passed back in full. This is a real integration constraint, not a footnote — it shapes how you build against the API.
• It acts on your behalf: Moonshot notes that when K3 "encounters minor issues, it may make unexpected decisions on the user's behalf." Budget for supervision on consequential tasks.
• Acknowledged user-experience gap: Moonshot states plainly that K3 "exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol." Benchmark parity is not product parity.
• Reasoning trails the frontier: HLE-Full at 43.5 versus Fable 5's 53.3 is a meaningful deficit on hard reasoning.
• Chinese data law: Kimi.com and the Moonshot API route prompts to Chinese servers, subject to China's Cybersecurity Law and Data Security Law. Use the open weights or a third-party host for sensitive data.
• Serving cost is real: 2.8 trillion parameters is not a laptop model. Sparse activation lowers inference cost, but the weights still have to live somewhere.
• Weights are promised, not shipped: As of launch, the July 27 date is a commitment. Open-weights plans have slipped before.

Best Use Cases

When to choose alternatives:

• Hardest reasoning tasks → Claude Fable 5 or GPT-5.6 Sol
• Polished end-user experience → Claude Fable 5 or GPT-5.6 Sol, per Moonshot's own assessment
• Agentic tool orchestration → Claude Fable 5 leads Toolathlon and APEX-Agents
• Smaller open-weights footprint → DeepSeek V4 Flash or Qwen 3.6

Getting Started

1. Try it free on kimi.com — the consumer chatbot defaults to K3
2. API access via the Kimi API Platform — note the cache-hit versus cache-miss input pricing split when you budget
3. Desktop agent via Kimi Work for Windows and Apple silicon Macs — research, visualization, and document creation
4. Terminal workflow via Kimi Code — select K3 with the /model command
5. Open weights — due by July 27, 2026; confirm the license before you build a distribution plan on it

Key Takeaways

• Kimi K3 is Moonshot AI's flagship — a 2.8 trillion parameter mixture-of-experts model activating 16 of 896 experts per token, with a 1 million token context window and native vision
• It is the largest open-weights model any Chinese lab has produced, and the first to land credibly inside the frontier conversation rather than a generation behind it
• K3 beats Claude Opus 4.8 on most published coding and agentic benchmarks — most dramatically on FrontierSWE, 81.2 against 66.7 — and leads its cohort on SWE Marathon, BrowseComp, and OmniDocBench
• Moonshot concedes K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, and names user experience rather than capability as the remaining gap
• Reasoning is the clearest deficit: HLE-Full at 43.5 against Fable 5's 53.3
• Every published number is vendor-measured; independent evaluation placed K3 third on ArtificialAnalysis at launch
• Output is priced at $15 per million tokens — frontier pricing, not a discount play; the aggressive rate is 30 cents per million cached input tokens
• Full weights are due by July 27, 2026, but Moonshot has not named a license — the K2 family used a Modified MIT License
• K3 requires preserved thinking history to be passed back in full, or generation quality can become highly unstable
• The launch landed one day before Xi Jinping made open source explicit Chinese national strategy at the World AI Conference in Shanghai