Moonshot’s Kimi K3: 2.8 Trillion Parameters, Open-Source, and Already Beating Anthropic on Coding
Moonshot AI’s Kimi K3 has 2.8 trillion parameters, tops Arena.ai’s frontend coding charts, and drops as open-source on July 27. China’s gap with the US just got a lot smaller.
Moonshot AI released Kimi K3 on July 16, 2026, and the AI world is still catching its breath. The Beijing-based lab describes it as its “most capable model to date, with 2.8 trillion parameters.” That is not a typo. At 2.8 trillion total parameters, K3 is roughly 75 percent larger than DeepSeek’s V4 Pro — the previous open-weight champion. Washington’s chip export controls were supposed to prevent exactly this kind of thing. Apparently no one told Moonshot.
The release was timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, which is either a coincidence or the most precisely planned PR move in Chinese tech history. Either way, it worked: chip stocks tumbled and Wall Street had a Friday it would rather forget.
What Kimi K3 Actually Is
Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model billed as the world’s first open 3T-class release, activating 16 of 896 experts per token, with a 1M-token context window and native vision. The sparse MoE architecture is what makes the scale tractable — the model doesn’t fire all 2.8 trillion parameters at once, just a focused 16-expert slice per token. It is built on two key architectural innovations: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which Moonshot describes as a drop-in replacement for residual connections that delivers consistent scaling gains. The company claims roughly a 2.5x improvement in scaling efficiency over Kimi K2, which is the kind of number that makes competing labs quietly schedule internal reviews.
Two variants shipped at launch: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing. On the API side, Kimi K3 is compatible with the OpenAI SDK, lowering the integration barrier for developers already building on OpenAI or Anthropic toolchains. Smooth move.
The Benchmarks: Strong, Not Flawless
On the Artificial Analysis Intelligence Index v4.1, K3 scores 57.1 — fourth overall, behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9), but ahead of Claude Opus 4.8 (55.7). The headline number that has developers genuinely excited, though, is the Arena.ai coding result. Arena placed K3 at number 1 on its Frontend Code Arena — a 17-place jump from Kimi K2.6’s position at 18 — with first place in six of seven frontend domains. In blind human-preference tests, developers picked K3’s frontend code over every other model on the board, including Fable 5.
On the agentic side, the numbers are equally impressive. On GDPval-AA v2, a benchmark covering real-world tasks across 44 occupations and 9 major industries, K3 scored 1,687 — third overall behind Fable 5 Max and GPT-5.6 Sol Max, but ahead of Claude Opus 4.8. On AA-Briefcase, testing long-horizon knowledge work, K3 climbed to second place, beating GPT-5.6 Sol Max. It scored 93.5% on GPQA Diamond and 91.2% on BrowseComp — the strongest open-weight results on those benchmarks at launch.
One caveat worth flagging: unlike every prior Kimi flagship, K3 was not open-weight on day one. Moonshot promises the full weights by July 27, but at launch there was no checkpoint, no licence, and no model card. Until those weights are public and independently reproducible, every benchmark number is a claim, not a proof.
Pricing: The Era of Cheap Chinese AI Ends Here
API pricing is $3 per million input tokens and $15 per million output tokens — putting it at the same level as Anthropic’s Claude Sonnet series and making it the most expensive model released by a Chinese AI lab to date. That is a significant increase on Kimi K2.6, which ran at $0.95 input and $4 output. The DeepSeek-style “frontier capability at bargain-bin prices” playbook is not entirely dead — Artificial Analysis estimates a cost per task of $0.94, close to GPT-5.6 Sol and about half of Claude Opus 4.8 — but Moonshot is clearly signalling that it considers K3 a premium product, not a price weapon.
The Geopolitical Read
Moonshot’s strategic pivot to open-source models — beginning with Kimi K2 in July 2025 and accelerating with K2.5 in January 2026 — was in large part an effort to reclaim relevance after DeepSeek’s rise knocked the company from third to seventh in monthly active users in China. K3 is the payoff on that bet. Moonshot’s own timeline chart places K3 as a dramatic outlier, towering above competitors like DeepSeek (1.6T), Xiaomi (1.02T), and Alibaba (397B).
Bank of America analysts noted that K3 shows large-scale pre-training plus architectural work can still deliver step-change gains for Chinese models despite compute constraints. That is the polite version. The blunt version: a Chinese lab just shipped the world’s largest open-weight model on hardware that was never supposed to be powerful enough to do it, and it beats most of what Silicon Valley is currently selling.
Moonshot’s $500M Series C, closed in January 2026 at a $4.3B valuation, was explicitly earmarked for K3 development and compute expansion. The company is also reportedly seeking a $30 billion valuation ahead of a planned Hong Kong listing — and K3 reads exactly like a document designed to justify that number.
What Happens on July 27
The weights drop on July 27, and that is when this story really starts. An API you can call is impressive; a 2.8-trillion-parameter model you can download, fine-tune, and run on your own infrastructure is a different category of disruption entirely. By releasing the world’s largest open-source model, Moonshot is making a bid to become the center of gravity for the global open-source AI developer community. If the weights hold up to independent scrutiny — and the architecture papers are already public on GitHub — the competitive pressure on closed frontier labs will be very real. OpenAI and Anthropic have been charging premium prices for models that, as of next week, a Chinese startup will be giving away. That is a conversation no one at those companies is excited to have.
