Kimi K3: Moonshot AI's 2.8-Trillion-Parameter Open Model Just Closed the Gap With Claude and GPT-5.5

What is Kimi K3?

Kimi K3 is the newest flagship large language model from Moonshot AI, a Beijing-based startup founded by former Google researcher Yang Zhilin. Released on July 16, 2026, it's a sparse Mixture-of-Experts (MoE) model built on 896 experts, with 16 activated per token, running on a new attention architecture Moonshot calls Kimi Delta Attention (KDA) paired with Attention Residuals (AttnRes). The headline number is size: 2.8 trillion total parameters, roughly 2.8 times larger than its predecessor, K2.6, and well above Chinese rivals like DeepSeek's V4 Pro (1.6 trillion) and Zhipu AI's GLM 5 series (744 billion).

The model supports a 1-million-token context window and native vision input, and Moonshot says it uses roughly 21% fewer output tokens than K2.6 to complete equivalent tasks — a meaningful efficiency gain given how output-token-heavy reasoning models have become.

How does Kimi K3 compare to Claude and GPT-5.5?

In its own release materials, Moonshot said K3 performs "competitively" with Anthropic's Fable 5 — currently the most capable publicly available model — and "substantially outperformed" Anthropic's Opus 4.8 and OpenAI's GPT-5.5 and GPT-5.6 Sol on the company's benchmark suite. Independent testing has been more mixed: some third-party evaluations, including one from Arena.AI, ranked K3 as the top-performing model currently available, while other testers flagged heavy reasoning-token consumption. One widely cited example: a simple SVG pelican-drawing test burned through more than 13,000 reasoning tokens, costing roughly $0.25 for a single query.

As with any lab-reported benchmark, K3's numbers should be read with the usual caveat — companies choose the tests that flatter their own models. But the direction of travel is clear: the performance gap between China's best open-weight models and the closed frontier systems from the US has been narrowing steadily since DeepSeek's R1 release in early 2025, and K3 is the latest and largest data point in that trend.

Pricing and availability

K3 is live now at kimi.com and through Moonshot's API, priced at $3 per million input tokens and $15 per million output tokens — the highest rate among major Chinese labs, but still roughly half the per-task cost of Anthropic's Opus 4.8, according to Moonshot's own comparisons. Cached input tokens are priced far lower, at $0.30 per million, which matters for coding workloads that repeatedly reuse large chunks of context. Full model weights are scheduled for release on July 27, 2026, under a Modified MIT license, making K3 the first fully open model in the 3-trillion-parameter class.

Why K3's release matters beyond the benchmarks

The timing is not incidental. K3 launched the same week China hosted the World AI Conference (WAIC) in Shanghai, where President Xi Jinping delivered his first-ever keynote at the event and called on the world to embrace open-source AI as a shared global resource rather than a tool controlled by any single country. Moonshot's stand was reportedly one of the most visited at the conference. K3's debut also rattled markets: chipmakers including TSMC and Nvidia saw shares dip amid renewed "DeepSeek shock" comparisons, as investors weighed whether cheaper, competitive open models reduce the case for continued massive spending on frontier compute.

For US policymakers, K3 adds to a growing body of evidence that export controls on advanced chips have slowed, but not stopped, Chinese labs from approaching the AI frontier — a point raised directly in a House Foreign Affairs Committee hearing on AI export controls just two days before the model's release.

What this means for developers and businesses

For teams evaluating AI infrastructure, K3 is a serious new option in the open-weight tier, particularly for long-context coding and agentic workloads where its 1-million-token window and lower cached-token pricing are meaningful advantages. The trade-off is heavier reasoning-token usage on some tasks, which can offset the headline pricing advantage depending on workload. As always with a same-week release, expect independent benchmark results to diverge from the lab's own numbers over the next few weeks — that's the number worth watching before making a production decision.

Is Kimi K3 free to use?

Kimi K3 is available through Moonshot's chat interface at kimi.com and via a paid API. It is not free for API access, but its per-task pricing is positioned below several closed competitors.

When will Kimi K3's full model weights be released?

Moonshot says full weights will be published on July 27, 2026, under a Modified MIT license.

Is Kimi K3 better than Claude or GPT-5.5?

On Moonshot's own benchmarks, K3 outperforms Opus 4.8 and GPT-5.5 and is competitive with Anthropic's more capable Fable 5 model. Independent, third-party verification is still emerging, so these claims should be treated as provisional.