On July 16, the same day the 2026 World Artificial Intelligence Conference opened in Shanghai, Moonshot AI rolled out Kimi K3. The first number the company put on the table was scale: 2.8 trillion parameters. Xinhua's same-night wire report called it "the largest-parameter open-source model in the world today," about 70% bigger than DeepSeek V4 Pro's roughly 1.6 trillion, and well ahead of earlier open models from Xiaomi (1.02 trillion) and Alibaba (397 billion).
Moonshot itself is calling K3 "the first open-source 3T-class model" — rounding 2.8 trillion up to 3 trillion. It's a Mixture-of-Experts (MoE) architecture: 896 experts in total, with only 16 activated per inference pass. It supports a 1-million-token context window, natively understands images, and ships with an always-on "thinking mode." Under the hood are two techniques Moonshot developed in-house and has already published: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which replaces standard residual connections. In the words of a Moonshot executive quoted by Xinhua, "nearly 3 trillion parameters means the model can pack more knowledge and patterns into its 'brain' — it knows more, thinks deeper, and answers more accurately."
Often first on individual tests, third overall
The real story with K3 isn't the parameter count — it's how close, and in some cases how far past, it pushes the two strongest closed models. On GDPval-AA v2, which covers real-world tasks across 44 occupations and nine industries, K3 places third, behind Claude Fable 5 Max and GPT-5.6 Sol Max but ahead of Claude Opus 4.8. On BrowseComp, a long-horizon information-retrieval benchmark, it posts a state-of-the-art 91.2 using a single-agent approach that runs the full 1-million-token context without compression. On Arena.ai's frontend-coding arena — a leaderboard ranked by human preference votes — it tops the field with a score of 1,679, putting clear daylight between itself and both Fable 5 and GPT-5.6 Sol.
Its coding-benchmark results look more like scattered wins: on SWE Marathon, a long-horizon coding test, it leads with 42.0 versus Opus 4.8 (40.0), GPT-5.6 Sol (39.0), and Fable 5 (35.0); on Program Bench it narrowly edges out GPT-5.6 Sol, 77.8 to 77.6; but it falls back behind both closed models on DeepSWE and FrontierSWE. It also tops Automation Bench and SpreadsheetBench 2, while on JobBench, Fable 5 (57.4) pulls well ahead of it (52.9). On visual reasoning benchmarks CharXiv and Zerobench, and on the composite agent leaderboard AA-Briefcase, it holds a steady second place. As officechai put it, K3's strength isn't finishing first on every test — it's the sheer density of wins across coding, agentic, vision, and knowledge-work tasks, achieved at a fraction of the cost of its rivals.
One thing worth flagging for the record: the numbers for the same leaderboards don't match across sources. VentureBeat lists K3's GDPval-AA v2 score as 1,687, while officechai has it at 1,668; AA-Briefcase shows up as both 1,527 and 1,548 in different reports, and Simon Willison, citing an official Artificial Analysis tweet, gives yet another figure — 1,547. None of the three outlets explain the discrepancy, so the original reports from Moonshot and Artificial Analysis should be treated as authoritative.
What's available now is the API — "open source" is still a step away
It's worth being precise here: all of these benchmark results currently come from Moonshot's own evaluations, or from early joint testing with the analysis firm Artificial Analysis. K3 is only available today via API and the kimi.com web app; the full model weights won't be released until July 27. In other words, the community can't yet independently reproduce these results. VentureBeat put it bluntly: whether the scores "hold up under independent evaluation" once the weights are public is the real thing to watch next.
The most expensive model out of a Chinese lab
Pricing has shifted too. K3 runs $3 per million input tokens and $15 per million output tokens, with cached input as low as $0.30 per million — putting it right in the same tier as Anthropic's Claude Sonnet line. That's a sharp jump from the previous generation, K2.6, which was priced at $0.95/$4 — both Simon Willison and VentureBeat singled it out as "the most expensive model a Chinese AI lab has released to date."
But measured per task, the math doesn't necessarily look bad. According to Artificial Analysis data, K3 costs about $0.94 per task, close to GPT-5.6 Sol's $1.04, and only half of Opus 4.8's $1.80. officechai's breakdown gets more granular: on Kimi Code Bench V2, K3 running at full reasoning effort hits roughly 73% accuracy at about $3.5 per task, while Fable 5 reaches about 78% but costs nearly $9 per task; on that 91.2 SOTA BrowseComp score, K3 gets there for under $2 per task, while GPT-5.6 Sol and Opus 4.8 need $5 to $27 to hit their respective scores. Moonshot now runs three tiers side by side: the flagship K3 ($3/$15), the coding-focused K2.7 Code ($0.95/$4), and the general-purpose K2.6 ($0.95/$4) — the latter two start at a 256K-token context, while only K3 offers the full 1 million.
Two demos designed to prove a point
Beyond the benchmarks, Moonshot released two demonstrations. In one, K3 was tasked with designing a physical chip capable of running a nanoscale version of itself: after 48 straight hours of autonomous agent operation using open-source EDA tools, running the entire pipeline from architecture design through optimization and verification, it produced a 4-square-millimeter chip that achieved timing closure at 100MHz, with simulated decoding speeds exceeding 8,700 tokens per second. The other involved the "I-Love-Q relation" in computational astrophysics — K3 reportedly reproduced in about two hours a calculation that would typically take a senior researcher one to two weeks, reading and cross-checking more than 20 papers along the way and implementing the complete numerical pipeline. Both cases come from Moonshot's own technical materials; no third party has reproduced either result.
Simon Willison took a more grounded approach. He ran his classic "generate an SVG of a pelican riding a bicycle" test on K3: 95 input tokens, 16,658 output tokens (13,241 of them reasoning tokens), at a cost of 25 cents for a single run. The expense comes down to the fact that K3 currently only offers a "max" reasoning-effort setting, with no way to dial it down. He also spotted a telling detail: even sending just "hi" burned through 86 tokens, suggesting the model carries a hidden system prompt of roughly 85 tokens — content the model declines to reveal. Willison himself cautions that this 21-month-old test has long since decoupled from actual model capability — GPT-5.6 and Fable 5 both draw worse pelicans than GLM-5.2, yet no one would call GLM-5.2 Fable-class. These days, he says, the test mostly just confirms whether a model can run at all.
The leak before the launch
The official announcement itself came off a bit clumsily. On the night of July 14, a screenshot began circulating in AI communities showing a "Kimi K3 launch limited-time top-up promotion" page, advertising tiered bonus vouchers for top-ups made between July 15 and August 11; the page was pulled within a few dozen minutes, but not before Google had already crawled it. The community's read is that someone on the operations team fat-fingered a status field to "published" while configuring the promotion. Around the same time, an anonymous model codenamed "Kivine" showed up on the Arena testing platform, turning in standout performance on web and frontend generation — widely suspected to be a hidden test version of K3, though Moonshot never confirmed it. The official launch two days later lined up neatly with these leaks. (One side note: the two versions of the top-up promotion don't quite match — VentureBeat describes a single top-up of $1,000+ earning up to 30% back, running through August 12, while the version circulating in Chinese-language communities describes tiered bonuses starting at 99 yuan, running July 15 through August 11. These may be two separate promotions — one for the international API, one for the domestic platform — but that hasn't been confirmed.)
A brief introduction to Moonshot AI
Moonshot AI was founded in 2023 by Yang Zhilin, a Tsinghua graduate who previously did research stints at Google and Meta. The company's fundraising has moved fast: by early 2026 it had raised roughly $1.5 billion, with its valuation climbing from $2.5 billion to $4.3 billion, and reports suggest it's now in talks for a new round at a $5 billion valuation.
Its product Kimi was once one of the leading AI assistants in China, but the January 2025 launch of DeepSeek R1 upended the entire Chinese AI landscape — Kimi's monthly active users fell from third place domestically to seventh, making it one of the hardest-hit players. The company responded by pivoting toward open source: K2 in July 2025, K2.5 in January 2026, and now K3. VentureBeat sees this as the company's most pivotal comeback moment in 18 months.