Moonshot AI released Kimi K3, a MoE model with 2.8 trillion total parameters and 50 billion active, the largest open-weight model ever published. Licensed under Apache 2.0, it’s available for self-hosting and commercial use. In early independent tests of its resistance to prompt injection, the model responded with a crisp «Is there something I can actually help you with today?» instead of breaking.

Reported performance puts it within reach of GPT-5.6 Sol and Claude Fable 5 on SWE-bench and complex coding tasks, and API pricing is no longer rock-bottom: input costs the same as Meta’s top tier (Muse Spark), output double. It’s a signal: China exits the race-to-the-bottom pricing tier and competes on value, not just cost.

Why this matters to you. If you work with self-hosted models, K3 is the first open one with frontier-level performance you can run on your own infrastructure. If you use APIs, pricing signals that the Chinese market is starting to position itself as a serious alternative to US labs, not just a low-cost mass supplier. Technical choices (MoE architecture, long context) mirror those of the latest proprietary models, and the Apache release pushes competitive pressure on other labs to open their weights.

In detail

Context.

Until now, open-weight models over a trillion parameters were research experiments (Grok 1, released by xAI in March 2024, had 314 billion) or distilled variants of larger models. K3 is the first model at this scale released as a complete product with a permissive license.

Moonshot AI is the company behind the Kimi assistant, popular in China for its long context window (originally 200k tokens, later extended). K3 is the third generation of the family, after K1 (2023) and K2 (early 2025). The model name follows the company’s internal naming convention.

The numbers.

K3 uses a Mixture of Experts (MoE) architecture: 2.8T total parameters, but only 50B active during inference. This lowers memory requirements and latency compared to a dense model of the same nominal capacity. Native context is 1M tokens, in line with recent frontier models (Gemini 2.5, GPT-5.6).

Reported benchmarks from Moonshot place K3 above Claude Opus 4.8 on AIME (math) and on par with GPT-5.6 Sol on SWE-bench Verified (real-world coding). Independent large-scale evaluations aren’t available yet, but early community testing confirms the model doesn’t break down on complex tasks.

API pricing is $1.50/million tokens for input and $6/million for output (verified from official pricing on July 17, 2026). For comparison: GPT-5.6 Sol costs $3/$15, Claude Fable 5 $3/$15, Muse Spark 1.1 $1.50/$3. K3 sits in the middle: input priced like Meta, output double. It’s no longer the Chinese pricing everyone knew (DeepSeek v3 was $0.14/$0.28), but high-tier model pricing.

What changes.

First, the quality gap closes. Chinese models are no longer “almost as good as” frontier US ones: on certain tasks they match them, and they do it with open licenses. This shifts the calculus for model selection: if K3 works for your use case, you have a self-hosted alternative that doesn’t depend on a foreign API.

Second, the pricing war changes tone. China has used price as a competitive lever for years (near-free open models via API). K3 signals that at least Moonshot now aims to compete on perceived quality, not dumping. If other Chinese labs follow, the global market rebalances: less race to the bottom, more differentiation on performance and licensing.

Third, pressure on other labs. Anthropic and OpenAI keep weights closed; Meta releases open but smaller (Llama 4 tops out at 405B dense, Muse Spark at ~600B MoE). An open model with 2.8T parameters, Apache licensed, that performs like their top tier, is hard for anyone asking for transparency to ignore.

The limits.

There are no independent large-scale evaluations yet: Moonshot’s benchmarks are self-reported. The community is testing the model, but it’ll take weeks for consensus on where it excels and where it stumbles. The model isn’t multimodal (text only), unlike GPT-5.6 and Gemini 2.5.

Self-hosting requirements are steep even with MoE: 50B active parameters need at least one machine with 4×A100 or H100 for smooth inference, and setup costs aren’t trivial if you don’t already have infrastructure. The API is the practical path for most use cases.

Finally, the ecosystem: tools and integrations for K3 are still immature compared to GPT or Claude. If your workflow relies on well-proven tooling, switching to a new model takes adaptation work.