Moonshot AI released Kimi K3 2.8T-A50B, the largest open model ever, claiming Opus 4.8-class performance at Sonnet 5 pricing. This marks a significant milestone in open-source AI, continuing a trend of major open model releases.
SynthePulse Insight · AI deep reading
Version 4 · 6 sources
Moonshot AI releases Kimi K3, a 2.8T-parameter open-source model that approaches the performance of GPT-5.6 Sol and Claude Fable 5, but its steep pricing and high deployment barriers mean the real test for the open-source community is just beginning.
On July 16, 2026, Moonshot AI officially released Kimi K3, a model with 2.8T parameters, which multiple media outlets have called the 'largest open-source model ever.' The model supports a 1M token context and has native multimodal capabilities.
Moonshot AI positions it as the 'world's first 3T-level open model' and promises to release the full weights before July 27.
According to official self-reported benchmarks, Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 on most tests but trails Claude Fable 5 and GPT-5.6 Sol, with the company describing the gap as a 'half-step.'
In the third-party Frontend Code Arena, Kimi K3 ranks first with a score of 1679, surpassing Claude Fable 5's 1631 and GPT-5.6 Sol's 1618.
Artificial Analysis evaluations show that Kimi K3 achieves an Elo of 1547 on long-horizon knowledge work assessments, a 732-point improvement over K2.6, second only to Claude Fable 5.
Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, matching Anthropic's Claude Sonnet series and becoming Moonshot AI's most expensive model to date, significantly higher than the previous K2.6's $0.95/$4.
Simon Willison's test showed that generating an SVG pelican image consumed 95 input tokens and 16,658 output tokens (including 13,241 reasoning tokens), costing 25 cents in total, highlighting its high reasoning costs.
Despite being open-source, Kimi K3's 2.8T parameter scale poses significant deployment challenges. Simon Willison notes that the model requires substantial GPU resources (e.g., 64+ accelerators), making it difficult for average developers to run locally.
Open-sourcing may bring risks of misuse, and there are questions about whether the community can effectively utilize such a large model. Additionally, it remains unclear whether Moonshot AI will offer quantized versions or an API.
Kimi K3's pricing strategy has been interpreted as 'the end of China's ultra-cheap AI era,' but this assessment is based on official pricing, and actual market reactions remain to be seen.
The specific quantitative criteria for the 'half-step' gap have not been fully disclosed, and performance comparison data partly comes from official sources and partly from third-party verification, introducing uncertainty.
This report is based on cross-verification from multiple sources. Facts such as model release, parameter scale, and pricing are confirmed; performance comparisons and the 'half-step' claim are primarily from official statements and are attributed as such; deployment barriers and open-source impact are inferences.
Kimi K3 sets a new record for open-source model scale with 2.8T parameters, approaching the performance of top closed-source models, but its high pricing and deployment barriers may limit its real-world impact, making adoption by the open-source community a key test.
Primary report
Primary source
Also covered by 5 sources