July 22, 2026 — On July 16, Moonshot AI dropped what may be the most significant AI model launch of the year. Kimi K3 is not just another incremental upgrade. At 2.8 trillion parameters, it is the largest open-weight model ever announced, and independent benchmarks place it fourth globally — within striking distance of OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 [citation:1][citation:4].

What makes K3 remarkable is not just its size. It is the first Chinese model to credibly challenge the American frontier on capability rather than price alone. And it is doing so with its weights set to go fully open on July 27 — a move that could reshape how enterprises and developers choose their AI infrastructure [citation:3][citation:5].

Core Specs: What Makes K3 Different

K3's headline numbers are striking, but the architecture behind them is what matters:

The 2.8 trillion parameter count is the headline, but the practical number is the active compute. With only 16 of 896 experts activated per token, K3 delivers frontier-level performance without requiring frontier-level inference costs [citation:2][citation:5].

Architecture Deep Dive: Delta Attention & MoE

Moonshot developed two in-house architectural innovations for K3:

Kimi Delta Attention (KDA)

A hybrid linear-attention mechanism designed to solve the long-context bottleneck. Standard transformers slow dramatically as context length grows. KDA reportedly delivers 6.3x faster decoding at 1 million tokens, making the full context window usable in practice rather than just a spec-sheet number [citation:5][citation:6].

Attention Residuals

Described as a drop-in replacement for standard residual connections. Moonshot claims a 25% training-efficiency gain for under 2% additional compute overhead — a critical advantage when training a 2.8T model under China's compute constraints [citation:5].

Stable LatentMoE

K3's mixture-of-experts framework activates just 16 of 896 experts per forward pass. This sparse design keeps inference costs manageable despite the massive parameter count, converting what would be an impractical model into a deployable one [citation:2][citation:6].

Benchmarks: How K3 Stacks Up

Independent testing gives a clearer picture than launch-day marketing:

The takeaway: K3 does not top every leaderboard, but it is competitive everywhere that matters. It leads on frontend coding, holds its own on general reasoning, and delivers the best long-context performance of any open-weight model [citation:4][citation:6][citation:7].

Head-to-Head: K3 vs GPT-5.6 vs Claude Fable 5 vs GLM 5.2

Against GLM 5.2 — Zhipu AI's coding-focused flagship — the comparison is different:

Pricing & Access: What It Costs

K3 launched with full availability across Moonshot's product suite:

• Kimi app & kimi.com — Free tier for everyone (iOS, Android, HarmonyOS, web)
• Kimi Work — Desktop app for Windows and Apple Silicon
• Kimi API — OpenAI-compatible endpoints at platform.moonshot.ai
• Kimi Code — Terminal/IDE agent with K3 selectable via /model

Moonshot also launched a referral program: new users who sign up through an invite link receive bonus membership credits — up to a full year free [citation:1].

Open Weights: The July 27 Milestone

The most consequential date on the calendar is July 27, 2026. That is when Moonshot has committed to releasing K3's full weights under a Modified MIT license [citation:1][citation:5].

If that happens, K3 will become:

• The largest open-weight model ever released — nearly 3x the size of K2's 1.0T
• The first open model in the 3T-parameter class
• A viable self-hosting option for enterprises with sufficient GPU infrastructure

Why K3 Matters for the AI Industry

K3's launch has already sent ripples through the market:

• Chip stocks wobbled — Investors fear open Chinese models could accelerate AI commoditization and reduce demand for high-end training chips [citation:6]
• Enterprise negotiating power shifts — A frontier-capable open model gives companies leverage against OpenAI and Anthropic pricing [citation:4]
• China's AI credibility rises — K3 is the first Chinese model to compete on capability, not just cost [citation:3][citation:6]
• Open vs closed debate intensifies — If a 2.8T open model can match closed frontiers, the case for proprietary lock-in weakens [citation:5]

Moonshot is also reportedly preparing for a Hong Kong IPO, with a valuation of approximately $31.5 billion. If successful, it would be the first Chinese large-model unicorn to go public [citation:6].

Key Takeaways