July 17, 2026. Moonshot AI released Kimi K3 on July 16, a 2.8 trillion parameter sparse mixture-of-experts model with native vision and a 1 million token context window. Moonshot calls it the first open 3T-class model, and VentureBeat reports it now sits fourth of 189 models on the Artificial Analysis Intelligence Index at a score of 57, behind only Claude Fable 5 and two GPT-5.6 Sol reasoning settings.
The key developments
- Architecture, not just scale. K3 activates 16 of 896 experts per token and introduces Kimi Delta Attention, which Moonshot says decodes up to 6.3x faster in million-token contexts, plus Attention Residuals, claimed to add roughly 25 percent training efficiency. Moonshot reports about 2.5x better overall scaling efficiency than Kimi K2.
- Benchmarks that trade blows with the frontier. In Moonshot's published suite, K3 leads on several agentic tests including BrowseComp at 91.2 and SWE Marathon at 42.0, while trailing Claude Fable 5 on FrontierSWE, 81.2 versus 86.6, and on HLE-Full. Moonshot itself states overall performance still trails Fable 5 and GPT-5.6 Sol. Vendor-published numbers, read them that way.
- Aggressive flat pricing. The API charges $0.30 per million tokens for cache-hit input, $3.00 for cache-miss, and $15.00 for output, with no tiering by context length and an OpenAI-compatible SDK. Moonshot reports cache-hit rates above 90 percent in coding workloads, which is where the economics get interesting.
- Open, with an asterisk for now. The weights are announced as open but were not yet downloadable at publication. VentureBeat reports the checkpoint is scheduled for release around July 27, and no K3 model card had appeared on Moonshot's Hugging Face organization as of July 17. Until the weights land, K3 is a cheap API model with an open promise, not yet a self-hostable asset.
What it means for operators
This extends the cost story we flagged when Chinese models started winning US workloads on price: the gap between frontier quality and commodity price keeps narrowing from below. For businesses running AI in production, the move is not to switch everything to the newest model. It is to re-run your routing math. Keep your highest-stakes reasoning on the models that lead your specific tasks, and test whether high-volume workloads, support drafting, classification, extraction, long-document processing, hold quality at a fraction of the cost on challengers like K3. The 1M context plus flat pricing makes long-context work, entire codebases or contract stacks in one call, newly affordable. Data residency still matters: K3 currently runs on Moonshot's infrastructure, so regulated workloads should wait for the weights and a US-hosted provider before routing anything sensitive. That evaluation, task-by-task benchmarking plus a routing layer, is exactly what our AI automation team builds, and our AI engineers can pressure-test a challenger model against your current stack in a week.
Frequently Asked Questions
Kimi K3 is a 2.8 trillion parameter sparse mixture-of-experts model from Beijing-based Moonshot AI, released July 16, 2026. It activates 16 of 896 experts per token, handles text, images, and video natively, and offers a 1 million token context window. Moonshot calls it the first open 3T-class model.
Partially, for now. The model launched on Kimi.com and the API on July 16, and Moonshot announced open weights, but the downloadable checkpoint was reported as scheduled for around July 27 and had not appeared on Hugging Face as of July 17, 2026. Until then it is an API model with an open-weights commitment.
On the Artificial Analysis Intelligence Index it scores 57, fourth of 189 models, behind Claude Fable 5 and two GPT-5.6 Sol settings. In Moonshot's own benchmark suite it leads on several agentic tests like BrowseComp and SWE Marathon while trailing Fable 5 on FrontierSWE and HLE-Full. Moonshot itself acknowledges the strongest proprietary models remain ahead overall.
API pricing is flat regardless of context length: $0.30 per million input tokens on cache hits, $3.00 per million on cache misses, and $15.00 per million output tokens, accessed through an OpenAI-compatible SDK. Moonshot reports above 90 percent cache-hit rates in coding workloads, which substantially lowers effective input cost for iterative work.