On July 16, 2026, Beijing-based Moonshot AI released Kimi K3, a mixture-of-experts language model with 2.8 trillion total parameters. Moonshot plans to publish the full model weights by July 27, making K3 the largest open-weight AI model released to date — well ahead of DeepSeek’s V4 Pro at roughly 1.6 trillion parameters, Xiaomi’s 1.02-trillion model, and Alibaba’s 397-billion-parameter system.
What Exactly Is Kimi K3?
K3 uses a mixture-of-experts architecture, meaning only a fraction of its parameters activate for any given request — roughly 32 billion active parameters per token, according to independent technical breakdowns. That design lets Moonshot claim a huge total parameter count while keeping inference costs manageable. The model also ships with native multimodal understanding, a one-million-token context window, and reasoning switched on by default, which Moonshot calls its “thinking mode.”
The Numbers Behind the Launch
- Released via API and Kimi’s apps on July 16, 2026; full open weights expected by July 27 under a Modified MIT license
- Pricing set at $3 per million input tokens and $15 per million output tokens, with cached input as low as $0.30 per million tokens
- Built on Moonshot’s Mooncake serving infrastructure, which separates prefill and decode processing to improve efficiency on coding workloads
How Kimi K3 Stacks Up
Early independent evaluations, including Artificial Analysis’s Intelligence Index and the Vals AI leaderboard, place K3 near the top of the field — generally trailing Claude Fable 5 and GPT-5.6 Sol, but ahead of Claude Opus 4.8 and GPT-5.5 on a number of coding and agentic tasks. One analysis even ranked it first on the Frontend Code Arena benchmark.
That said, most of these figures are still based on Moonshot’s own reporting or limited API access. Until the weights are actually public on July 27, outside researchers can’t fully reproduce the results, so the rankings should be treated as provisional rather than settled.
Why This Launch Is Bigger Than One Model
Giving away a frontier-scale model’s weights is a deliberate strategy. It lets Moonshot’s technology spread across the developer community far beyond what its own infrastructure could serve directly, building influence the way Meta’s Llama models once did for the open-source ecosystem. It’s also reignited a broader industry argument: some commentators have warned that a world of powerful open-weight models is inherently harder to control, while others see K3 as healthy evidence that the gap between closed frontier labs and the open community is narrowing.
There’s a more pointed controversy attached to this release, too. Anthropic previously said Moonshot used millions of logged Claude conversations to help train an earlier Kimi model through distillation. K3’s benchmark scores now land close to the very models named in that dispute, which has led some observers to ask whether a similar pattern is at play here. Moonshot hasn’t publicly addressed the comparison.
The launch has also drawn a policy response. Reports indicate that U.S. officials are reconsidering restrictions specifically targeting Chinese-developed AI models, citing cybersecurity concerns raised after K3’s release — part of a wider pattern of AI policy shifting from voluntary guidelines toward more direct government intervention.
What It Means If You’re Choosing AI Tools
For teams evaluating models for coding, agentic workflows, or high-volume tasks, K3’s combination of strong benchmark performance and low per-token pricing makes it worth testing, particularly through an API or inference provider rather than self-hosting. Running the full model yourself is a serious undertaking: Moonshot recommends serving it on “supernodes” of 64 or more accelerators working together, which puts self-hosting out of reach for all but the largest infrastructure teams. Most developers will access K3 through Moonshot’s own API or third-party inference platforms instead.
For the highest-stakes or most sensitive work, the top-ranked closed models still appear to hold a narrow edge in these early results — so K3 looks best suited as a cheaper, capable option for the bulk of everyday development work rather than a wholesale replacement.
What to Watch Next
The real test arrives on July 27, when Moonshot is scheduled to publish K3’s full weights. That’s when independent researchers can finally verify whether the benchmark claims hold up outside Moonshot’s own testing environment, and whether the model can be adapted for the specialized, domain-specific uses that make open-weight releases valuable in the first place.
