Kimi K3 (Free)

This August 7, 2026 head-to-head from Qubrid AI compares Kimi K3 against Alibaba's Qwen3.8-Max, both frontier-class open-weight MoE models with 1-million-token context windows, native multimodality, and significantly lower pricing than leading proprietary alternatives. Kimi K3 was released through Moonshot's API on Jul The supplied comparison breaks down workload-specific strengths: K3 leads on terminal and repo-scale software engineering (88.3 Terminal-Bench 2.1 vs 86.6; 81.2 FrontierSWE vs 73.5), while Qwen3.8-Max leads on agentic computer use (86.1 OSWorld-Verified vs 84.8). Token pricing on Qubrid is $3.00/$15.00 per million for

SheetsX's August 5, 2026 reference page describes Kimi K3 as Moonshot AI's most capable model as of August 2026, combining a sparse Mixture-of-Experts architecture, native visual understanding, persistent reasoning, and a 1-million-token context window sufficient for extensive repositories or document collections. The The pricing table from the official Kimi API Platform dated August 5, 2026 lists kimi-k3 at $0.30 cached input, $3.00 uncached input, and $15.00 output per million tokens with a 1,048,576 token context window. The supplied excerpt reveals a technically useful API detail: K3 always reasons, and the API exposes low, high

A July 30, 2026 Rest of World article reports that Moonshot AI released the weights behind Kimi K3 openly on July 27, allowing governments, companies, and individuals to run and fine-tune the model locally without paying Moonshot, framing it as a potential shift in sovereign-AI economics. The article notes Kimi K3 rank The same piece contextualizes Kimi K3's open-weight release as a response to nations spending billions on sovereign AI infrastructure while leasing software from U.S. cloud providers, quoting a Middle East Institute senior fellow on how a competitive free model changes the ROI calculation for government hardware invest

Sid Saladi's July 22, 2026 guide reports that Kimi K3 scored 90.49/100 on the author's proprietary 16-task blind benchmark, described as beating every Claude and GPT model tested. The benchmark consists of the same 16 real-work tasks used in prior model comparisons, spanning engineer, founder, PM, and generalist catego The Substack guide serves as a Kimi K3 101 covering what the model is, the benchmark numbers, every access path including a $0 route, real cost math against Claude and GPT, and copy-paste prompts built around what K3 specifically does well. While full methodology details are paywalled, the supplied excerpt establishes

Kimi K3 is Moonshot AI's flagship multimodal reasoning model, launched on July 16, 2026, with 2.8 trillion total parameters in a sparse Mixture-of-Experts architecture that activates only 16 of 896 routed experts per token, making it the first announced open model in the three-trillion-parameter class. It ships with a At launch, K3 ranked first in Arena's blind frontend coding ranking and posted competitive results across Moonshot's coding and agentic benchmark suite, though the excerpt notes most detailed scores are vendor-reported and use mixed agent harnesses. Pricing is set at $3 per million uncached input tokens, $0.30 per mill

Puter developer Reynaldi Chernando published a review on July 20, 2026 describing Kimi K3 as Moonshot AI's new flagship released July 16, 2026, with roughly 2.8 trillion parameters in a Mixture-of-Experts architecture, positioning it as the largest open-weight model announced to date ahead of DeepSeek V4 Pro at 1.6 tri The review includes a hands-on visual-coding test of K3's claim to use visual feedback for frontend work, reports K3's ability to see images (confirming multimodal input), and compares pricing relative to Kimi K2.6. Puter is a developer/inference platform adding K3 to Puter.js, so the piece carries a mild commercial fr

Build Fast with AI's July 17, 2026 review notes Kimi K3 landed on July 16, 2026, as the largest open-weight model ever announced, at $3 input and $15 output per million tokens, representing a roughly 5x price increase over its predecessor. The review positions K3's release as Moonshot abandoning the cheap-alternative s The review identifies three defining design choices: scale (2.8T total parameters nearly triples the 1T blueprint the K2 family shared), multimodality (K3 reasons over images and video natively, including frame-by-frame video questions, unlike the text-first K2 line), and always-on thinking with no non-reasoning mode,

This July 20, 2026 review from Tabbit describes Kimi K3 as a Fable 5-level model priced at Sonnet 5 rates, noting independent benchmark numbers place it close behind Fable 5 and GPT-5.6 Sol rather than ahead, which still makes it the highest-ranked open-weight model announced to date. The review highlights two practica Rather than re-reporting standard benchmarks, the Tabbit review focused its original testing on K3's ability to use visual feedback for frontend work. The review notes Moonshot reports coding benchmark results on par with Claude Fable 5 and GPT-5.6 Sol, models costing roughly 1.7x to 3.3x more per token. Specifications