Chinese AI lab Moonshot AI published the full open weights of Kimi K3 on July 27, 2026, a 2.8-trillion-parameter model that, by total parameter count, is the largest open-weight release any AI lab has ever put out. The model itself, accessible via API, first became available on July 17; the open-weights drop ten days later is the part that actually changes who can use it and how.
K3 is a mixture-of-experts model, meaning that despite its enormous total parameter count, only a fraction of the network activates for any given token. In K3’s case, just 16 of its 896 experts fire per token, working out to roughly 50 billion active parameters — a figure that puts its actual per-token compute cost much closer to a mid-sized model than a 2.8-trillion-parameter one might suggest. The full weights, published in MXFP4 four-bit precision, still add up to roughly 1.4 terabytes, and the model supports a 1-million-token context window.
That footprint is the real story for anyone hoping to run K3 themselves. This is not a model that fits on a single high-end consumer card, or even a well-specced workstation. Loading the full weights requires a multi-accelerator server node — Nvidia Blackwell or AMD MI400-class hardware — and even an eight-card node with 192 gigabytes of memory per card is described as barely accommodating the model, with little headroom left for context or handling more than one request at a time. A single RTX 4090 or a Mac Studio, the kind of hardware a well-funded startup or research team might actually own, cannot run the full model even with aggressive quantization. In practical terms, open weights here means open to companies and universities with real GPU infrastructure, not open to an individual developer on a laptop.
On the license question, Moonshot’s own launch materials reportedly did not include finalized terms for the weights as of this writing, which is worth flagging plainly rather than assuming away: the company’s prior open release, Kimi K2.7 Code, shipped under what it called a Modified MIT license — MIT terms plus a single attribution clause that adds extra obligations once a product built on the model crosses 100 million monthly active users or $20 million in revenue. Kimi K3 is widely expected to follow the same template, but that had not been officially confirmed at publication, and any team planning a commercial product on top of K3 should check the actual license file once it lands rather than assume the terms carried over automatically.
On capability, independent evaluations have placed K3 second on one general intelligence benchmark and third on another, behind Anthropic’s Fable and OpenAI’s GPT-5.6 Sol Max in both cases, while reportedly taking first place on a benchmark specifically measuring frontend coding output. That is a genuinely different competitive position than earlier open-weight Chinese models occupied: K3 is not winning by being cheap and merely adequate, it is landing within a few points of the most capable closed models in the world on general reasoning, while beating them outright in at least one practical coding category.
Moonshot has framed the release less as a product launch and more as a bet on a specific strategic question the industry has been circling for a year: whether the most capable AI models end up being things enterprises rent from a handful of API providers, or things they can actually own and run themselves. Making a frontier-capable model’s full weights genuinely downloadable, not a smaller distilled version but the actual model, is Moonshot’s answer, and it puts real pressure on the two biggest closed-model labs to justify why their best work should stay locked behind an API.
It also matters that this is happening at 2.8 trillion parameters and not at the smaller, more portable scale where open-weight releases have traditionally competed. For the past few years, the pattern among Chinese labs releasing open models was broadly the same: ship something efficient enough to run on modest hardware, trade some raw capability for accessibility, and compete on cost rather than the top of the leaderboard. K3 breaks that pattern by refusing to make the trade-off at all — it is both enormous and, per the third-party rankings above, genuinely close to frontier capability, which forces a different conversation than “good enough and cheap.” A lab willing to give away a model this large, with a license this permissive, is making a statement about where it thinks the real competitive moat sits, and it clearly is not believed to be the weights themselves.
What This Means for Philippine Founders
For virtually every Philippine startup, Kimi K3’s open weights land as an interesting industry signal rather than something to actually deploy. The hardware bar alone, an eight-GPU Blackwell or MI400-class server node, puts self-hosting well out of reach for anything short of a well-capitalized local cloud provider or a university research lab, and even those will want firm license terms in hand before committing infrastructure to it. The more immediate relevance is competitive: a near-frontier model with a genuinely open license, once confirmed, gives cloud infrastructure providers and API resellers, including any operating in Southeast Asia, a credible, cheaper option to license and re-host, which over time should put downward pressure on API pricing across the board, Anthropic and OpenAI included. For Philippine founders building on rented API access rather than owned infrastructure, that competitive pressure between open and closed frontier models is worth watching closely, since it is likely to show up as lower per-token costs on the platforms they are already using long before any local team is running Kimi K3 on its own servers. It is also a useful data point for any Philippine team weighing a build-versus-buy decision on AI infrastructure: the honest answer, for now, is still buy — but the gap between what you can rent and what a determined competitor could eventually own is visibly narrowing.
Share this article