The biggest open-weight model ever shipped went live today

At 00:00 UTC on July 27, 2026, China's Moonshot AI did exactly what it said it would: it published the complete weights for Kimi K3. Two-point-eight trillion parameters. Larger than any open-weight model released before it, and the first downloadable model competing in the same league as the top closed systems from American frontier labs.

Here's the deal, though — the story only makes sense when you put two numbers side by side.

The first is third place. On the Artificial Analysis Intelligence Index, Kimi K3 sits behind only Claude Fable 5 and GPT-5.6 Sol Max. The third-most-capable model on earth is now free to download.

The second is 1.4 terabytes. That's the total size of the weight files. After four-bit quantization.

It's free and you can't have it. More precisely: you can have it, and you have nowhere to put it. The distance between those two numbers is the entire state of the open-weight ecosystem in 2026.

Who Moonshot is, and where K3 came from

Moonshot AI (月之暗面) was founded in Beijing in March 2023. Founder Yang Zhilin holds a PhD from Carnegie Mellon and is a co-author on the Transformer-XL and XLNet papers. The company name comes from Pink Floyd's The Dark Side of the Moon, and Alibaba and Tencent are among its major backers.

Moonshot's thesis has been consistent from day one: long context. Its first product, Kimi Chat in 2023, made noise in China by advertising a 200,000-character context window, and every generation since — K1.5, K2 — has pushed on context length and agentic capability. Kimi K2 in 2025 was the inflection point: a trillion-parameter-class MoE model released with open weights, which is when Western developers started taking Chinese open models seriously as production options rather than curiosities.

K3 is the same thesis at a much larger scale. The model itself launched on July 16, available first through the API and Moonshot's own products — Kimi.com, Kimi Work, Kimi Code. The weights followed eleven days later, today. That sequencing isn't accidental. Monetize through your own API first, let community evaluation accumulate, then open the weights. Moonshot ran the same play with K2.

Within the Chinese open-weight field, K3 now sits at the top. DeepSeek, Alibaba's Qwen, and Zhipu's GLM series all keep shipping capable open models, and K3 currently leads them on raw capability. Calling it the second act of the shock DeepSeek R1 delivered in early 2025 isn't much of a stretch.

What's inside

K3's architecture takes the modern MoE philosophy — build it enormous, activate a sliver — about as far as anyone has taken it.

Item Spec
Total parameters 2.8 trillion
Architecture Sparse MoE — 16 of 896 experts active per token
Active parameters ~50 billion
MoE framework Stable LatentMoE
Key techniques Kimi Delta Attention (KDA), Attention Residuals (AttnRes)
Context window 1 million tokens
Quantization MXFP4 weights / MXFP8 activations
Total weight size ~1.4TB
Recommended deployment Supernode with 64+ accelerators
Scaling efficiency vs K2 ~2.5×
API pricing $0.30 input (cache hit) / $3.00 (cache miss), $15.00 output per MTok
Weight release July 27, 2026

The concept to understand is Mixture of Experts. Those 2.8 trillion parameters don't all fire on every token. The model contains 896 "expert" sub-networks, and for each token a router selects 16 of them. Roughly 50 billion parameters actually participate in the computation. Compute cost of a 50B model, knowledge capacity of a 2.8T one.

The catch is that MoE reduces compute cost, not memory cost. You can't know which experts a token will need until it arrives, so all 2.8 trillion parameters must sit resident and instantly reachable. Even squeezed to four bits, that's 1.4TB. Community analysis on Hugging Face puts a practical minimum at eight nodes of eight 80GB GPUs each — 5.12TB total. Moonshot itself recommends supernode configurations of 64 or more accelerators.

Two genuinely new pieces of engineering here: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Both target the way information degrades as it passes through stacked attention layers, and Moonshot credits the combination with roughly a 2.5× improvement in overall scaling efficiency versus K2. That's the machinery that makes a one-million-token context window actually usable rather than nominally supported.

On performance, Moonshot's own blog is refreshingly straight. It claims "frontier-level performance across our evaluation suite" while explicitly noting that K3 trails Claude Fable 5 and GPT-5.6 Sol. But in specific domains it clearly leads.

Benchmark Kimi K3 Note
Frontend Code Arena 1,679 (#1) Fable 5 at 1,631 / GPT-5.6 Sol 1,618 / GLM-5.2 1,587
SWE Marathon 42.0 Leads on coding
Program Bench 77.8 Leads on coding
DeepSWE 67.3 —
GPQA-Diamond 93.5 Graduate-level science reasoning
MMMU-Pro 81.6 Multimodal understanding
DeepSearchQA (F1) 95.0 Agentic search
BrowseComp 91.2 Web browsing agent
Artificial Analysis Index #3 overall Behind Fable 5 and GPT-5.6 Sol Max

Beating Claude Fable 5 outright on frontend code generation is the detail worth pausing on. Third overall, but first on one of the tasks people use these models for most.

What each side gets

Moonshot gets standard-setting status. Publishing weights doesn't generate revenue directly. What it generates is researchers worldwide adopting K3 as a baseline, fine-tuning on top of it, and listing it as the comparison row in their papers. That's precisely what DeepSeek harvested from R1. And the recognition converts into API revenue. K3's API runs $3 per million input tokens and $15 output — the highest of any Chinese lab. Open weights as the lure, premium pricing on the hook.

China's AI ecosystem gets a narrative. With US export controls restricting access to top-tier accelerators, reaching third place globally while spending, as Interconnects puts it, "orders of magnitude less capital" than American counterparts is a potent political message regardless of what you think of the benchmarks.

Enterprises get leverage and sovereignty. If you have the infrastructure, you can download K3 and self-host. Data never leaves your perimeter, cost becomes depreciation rather than per-token billing, and no vendor deprecation can kill your service. In finance, healthcare and defense — anywhere data egress is regulated — that outweighs a few points of benchmark difference.

Individual developers get… honestly, not much. A 1.4TB checkpoint is not something a person handles. What they do get is price pressure. When the best open model tracks three to five months behind the frontier, closed-model API pricing has to respect that. A meaningful part of why per-token costs have fallen over the past two years is simply that open models exist.

And frontier labs lose something, which is worth stating for fairness. As Interconnects argues, open models compress frontier-lab margins. Compressed margins mean less capital for the next training run, which can slow progress over the long run. Arguments that open weights are unambiguously good tend to skip past this.

How past releases have gone

The success case is Meta's Llama. Releasing Llama 2 in 2023 under commercially usable terms effectively started the open-weight ecosystem. Fine-tuning tooling, quantization formats, local inference engines — all of it grew on that foundation. Meta didn't get revenue; it got the standard. Through Llama 3 and 4 it lost the top-capability race, though, and practical leadership of the open-weight field has since moved to Chinese labs.

The second success is DeepSeek R1. In January 2025 a reasoning model trained at a claimed fraction of the usual cost shipped with open weights and genuinely moved US equity markets. Two lessons stuck. First: an open-weight release is a geopolitical event, not just a technical one. Second: when it ships in a form the community can verify, the blast radius is far larger and far faster.

The cautionary case is quieter. Large open models have been released before and then forgotten because nobody could run them. The file existed, and no fine-tunes, no evaluations, no derivatives followed. The real value of an open release scales with how many people can actually execute it, not with parameter count. K3 is squarely on that test. For most of the community 1.4TB is a wall, and practical usage will concentrate among infrastructure-rich enterprises and cloud providers who resell access.

One more piece of context. In the last few days Jensen Huang used his first-ever X post to amplify an open-weights letter, while Washington debates restrictions on Chinese open-weight models. K3's release lands in the middle of that fight. Fortune's coverage of Huang's letter mentioning Kimi in the same breath was not a coincidence.

How US labs counter

Price. The most direct lever. When open weights trail by three to five months, closed models can only charge the premium that gap justifies. The steady expansion of cheap tiers at OpenAI and Google is that pressure showing up in the pricing page.

Productization. Weights alone are not a service. Tool calling, code execution sandboxes, browsing, memory, safety filtering, enterprise administration — all of it is engineering outside the model. What Claude Code and ChatGPT actually sell is that entire shell, and downloading K3 does not come with one. This is precisely where frontier labs have concentrated their effort over the past year.

Policy. The strongest and most contentious card. The Trump administration has reportedly weighed Entity List designations, procurement pressure, and security-liability rules aimed at Chinese AI developers. No blanket ban has been announced, but measures that make it hard for US companies to build commercial services on Chinese open weights are plausible. The card cuts both ways, though: tighten too far and the American open-weight community gets caught in the same net, which is exactly why Nvidia, Meta and Microsoft signed the letter opposing it.

Inference infrastructure. Running K3 requires a supernode, and the companies that build supernodes are overwhelmingly American. AMD announcing Helios into full production — 31TB of HBM per rack — on the same day is worth reading in this light. A Chinese lab can open the model; the leverage over the hardware layer that runs it at scale still sits in the US.

What actually changes for you

If you're an individual developer, don't bother with the download link. Do try the API. Frontend code generation benchmarks at the very top, so running your existing prompts against K3 side by side with whatever you use now takes about thirty minutes. Check whether sending code to a Chinese-hosted service clears your org's policy first.

If you're a CTO, the decision isn't about capability, it's about total cost of ownership. Self-hosting K3 requires multiple terabytes of GPU memory plus the people to operate it. Against API pricing, the break-even only arrives at substantial volume. Unless data sovereignty is a hard regulatory requirement for you, the API is the rational choice at this stage.

If you run AI infrastructure, this is straightforwardly an opportunity. "K3 hosting" offerings will appear from clouds shortly, and most real consumption of open-weight models flows through exactly these intermediaries. The DeepSeek R1 cycle repeated this pattern precisely.

If you work on policy, today is an inflection point. "Chinese open-weight model" no longer describes a weaker alternative — it describes the world's third-ranked model. That changes the premise of the regulatory conversation entirely, and Washington's next move may land within the week.

If you're just a user, there's little to feel directly, but the direction is favorable. Every time a top-tier model goes open, pricing and free quotas on closed models feel the pressure. Expect this news to show up quietly in your chatbot's next pricing revision.

🥄 Three Things You're Probably Wondering

— Can I actually download and run this? Realistically, no. Even at four bits the weights are 1.4TB, and practical deployment starts around 64 80GB GPUs. That's out of reach not just for workstations but for most mid-sized company server rooms. As with Llama and DeepSeek, the community will likely produce more aggressive quantizations within days or weeks — but at this size, expecting a "runs on a laptop" build is unrealistic.

— Then why does this matter? Because what changes the market isn't whether you can download it — it's that it's downloadable at all. When the world's third-best model can be obtained for free, closed models can only charge for the gap above it. And any company or country with infrastructure can now reach frontier-class capability without a vendor relationship. You not being able to run it doesn't mean it doesn't affect you.

— Is it actually better than Claude or GPT? Not overall, no. Moonshot's own blog states plainly that K3 trails Claude Fable 5 and GPT-5.6 Sol, and the Artificial Analysis composite puts it third. It does take first place on specific tasks like Frontend Code Arena. "Slightly behind overall, ahead in particular domains" is the accurate summary, and the only question that matters is whether your workload lands in one of those domains.

References

Numbers and criteria are as of announcement and may change.