According to Financial Times report wey cite anonymous sources, Moonshot’s Kimi K2 models don perform well for open source AI market. Dem rank high for benchmarks and show capabilities wey no too far behind the latest frontier models.
Now, the company wey get name Moonshot don prepare to release Kimi K3, wey go be the biggest open-weight AI model from China. FT say the model go get parameter count between 2 trillion and 3 trillion, and dem go release am for the coming days.
Moonshot raise $2 billion at $20 billion valuation for May. The news come as people dey argue whether e dey worth am to pay AI labs like OpenAI and Anthropic for their expensive closed-source models.
Industry leaders dey fear say AI labs go fit extract the data wey clients submit for use with products like ChatGPT and Claude, and train their own models with am.
For Moonshot’s own benchmarks, Kimi K3 perform on par with leading proprietary models. The company call K3 the first open model wey dey roughly 3 trillion parameter range. Dem schedule full model weights release for July 27.
The model target long-running programming tasks, knowledge work, and complex reasoning. All results come from Kimi and dem achieve am at maximum or high thinking intensity.
Across all 35 tests, K3 take first place about seven times and land second or third for most of the rest. Fable 5 win the most individual tests.
Independent testing lab Artificial Analysis don publish first evaluation of Kimi K3. Their results largely line up with Kimi’s own claims.
K3 also take top spot on AutomationBench-AA, Artificial Analysis’s version of Zapier’s agentic SaaS workflow evaluation, with score of 53 percent. Only Claude Fable 5 score higher.
Artificial Analysis call K3 well-rounded, with rubric scoring and analytical quality close to Fable 5 level. Sol still lead on presentation quality.
K3 accuracy rate improve from 33 percent to 46 percent on the AA-Omniscience Index, pushing overall score from +6 to +18. But hallucination rate climb from 39 percent to 51 percent, meaning K3 fabricate more answers even as e get more questions correct.
According to Kimi, the model primary use case na long-running software development with minimal human oversight. K3 dey built to analyze large codebases, coordinate terminal tools, and stay focused on task across many work steps.
The model pair programming with visual feedback: e examine screen captures, modify code, then check visible output. Kimi call this closed-loop system ‘Vision in the Loop’ and position am as foundation for game development, UI design, and CAD.
K3 use mixture-of-experts architecture wey activate only 16 of 896 experts at a time. ‘Attention residuals’ reportedly boost training efficiency by about 25 percent while adding less than 2 percent in extra compute overhead.
Chinese providers no dey offer their frontier models at rock-bottom prices anymore. But K3 still cheaper than top Western models and sit for upper midrange. Anthropic new Sonnet 5, for example, cost $3 per million input tokens and $15 for output but deliver lower performance.
K3 also use fewer tokens than its predecessor. For OpenRouter, the model list under identifier ‘moonshotai/kimi-k3,’ though e currently serve only through Moonshot itself. The open weights expected by end of July.
