Kimi K3: A 2.8-Trillion-Parameter Open-Weight Model — How Far Have Open Weights Come?

Kimi K3 is the model Moonshot AI released in July 2026 — 2.8 trillion parameters, but more importantly, weights you can download. A free model now stands on the same field as the paid commercial flagships. The first question most people ask is whether it runs on their own machine. The short answer is no: even the most aggressive compression needs around 594 gigabytes of memory. Free to obtain and possible to run locally are two very different things — and that gap, along with a license that is not quite what it looks like, is what this article is about.

Two releases, eleven days apart

Moonshot announced Kimi K3 on July 16, 2026, and published the weights on Hugging Face on July 27 — cite carefully, because the two dates get conflated. With the weights out, Together AI and Modal offered day-one hosting and the model became reachable through OpenRouter, so an open release immediately meant competing providers rather than a single vendor.

How to read 2.8 trillion

The total is 2.8 trillion parameters, but the mixture-of-experts design activates only 104 billion per token. Think of the total as books on the shelf and the active count as books actually opened per question. Compute per inference scales with the active count, so the model runs lighter than the headline suggests — but since you cannot know in advance which experts will fire, the entire weight set must sit in memory. Light to run, brutal on memory: nearly everything about deployment follows from that one property. Context length is 1,048,576 tokens, about one million. Input is natively multimodal — text and images — while output is text only; it reads pictures, it does not draw them.

Where it lands on independent measurements

On the Artificial Analysis Intelligence Index, Kimi K3 scored 57 at release — third among all models on July 17, below Claude Fable 5 and GPT-5.6 Sol and level with Claude Opus 4.8 and GPT-5.5. After scoring updates it stands at 60 as of early August 2026, in fourth place since Claude Opus 5 (now 63) arrived a week later. Rankings move in weeks, not years. The same firm measured 1668 Elo on GDPval-AA v2, up sharply from the 1190 of the previous Kimi K2.6, and 1547 Elo on AA-Briefcase — while consuming 21 percent fewer tokens than its predecessor across the full evaluation suite. Smarter, and less long-winded about it.

First place where it counts: autonomous work

Narrow the lens and the picture changes: on AutomationBench-AA, which tests multi-step autonomous task completion, Kimi K3 took first place at 53 percent — ahead of every commercial flagship. Composite indexes reward one-shot depth on hard questions; autonomous work rewards recovering from failures, fetching missing information, and not losing the plot mid-procedure. Those are different muscles. If you plan to let a model drive multi-step workflows, a benchmark like this is closer to your use case than any overall rank.

Pricing: about a third of the top commercial tier

Against Claude Fable 5 it runs about a third the price on both sides; against Claude Opus 5 and GPT-5.6 Sol, roughly 60 percent. The 0.30-dollar cached-input rate is a 90 percent discount for anyone resending a fixed preamble. Measured cost per Intelligence Index task came to 0.94 dollars, against 2.03 for Claude Opus 5 and 1.04 for GPT-5.6 Sol — a few points less capability for meaningfully less money.

Open weight is not open source

This is the part to get exactly right. Kimi K3 ships under a custom Kimi K3 License, not a standard open-source identifier. Two revenue-linked conditions apply: running it as a hosted inference business requires a separate agreement with Moonshot once that revenue passes 20 million dollars in twelve months, and embedding it in a product with over 100 million monthly active users or 20 million dollars in monthly revenue triggers a duty to display the Kimi K3 name. Almost no reader will hit either bar — individuals, startups, and most companies may use, modify, fine-tune, and redistribute freely. But open-source licenses by definition do not discriminate by user scale, and this one does. Blur that distinction in an internal tech evaluation and legal will un-blur it for you later.

Will it run at home? The 594-gigabyte floor

The weights themselves are 1.56 terabytes. To run them, everything must fit in memory: the vLLM project’s metadata estimates a floor of about 1.68 terabytes of VRAM at full precision. Quantization — trading numeric precision for size — only goes so far: INT4 still needs about 1,515 gigabytes, and the most aggressive 1-bit distribution still requires 594 gigabytes, a 62 percent cut that remains twenty-odd times the memory of a high-end consumer GPU. Moonshot’s own deployment guidance assumes clusters of 64 or more accelerators. Renting cloud GPUs at this scale runs to tens of dollars per hour, at which point the API is usually cheaper. For what individuals can realistically run at home, see our local open-model guide — the ceiling is far below this.

Why the open weights still matter

If you cannot run it, does the release matter? Three ways, yes. Competition: multiple providers serve the same model, so you choose on price and latency, and a vendor outage or price hike has a structural exit — no small thing in a month when regulation visibly moved availability, as we covered in our dependency-risk piece. Confidentiality: organizations that cannot send data out but do own serious compute get flagship-class capability in-house, which no closed model can offer even in principle. And research: fine-tuning, quantization experiments, and interpretability work all start with weights in hand. Most individuals will touch Kimi K3 through an API — but the open weights are why that API has competitive prices and multiple sellers.

Among open-weight peers

Within the open-weight field, Kimi K3’s 60 clears GLM-5.2 at 51 (753 billion parameters) and DeepSeek V4 Pro at 44 (1.6 trillion) — a clear head, on the largest frame. Moonshot’s own model card claims 93.5 on GPQA Diamond, 67.5 on DeepSWE, and 88.3 on Terminal-Bench 2.1; as always, vendor-picked terrain, useful color, not a comparison basis. A reasoning_effort parameter offers low, high, and max, defaulting to max.

At the workbench

For makers: plan around the API, not local hardware. Two uses look genuinely practical. Image-in failure diagnosis — feed photos of failed prints and ask for candidate causes, or triage surface defects in bulk, workable at a 3-dollar input rate. And cached repetition: if every session starts with your machine profiles, filament data sheets, and failure log, the 90 percent cache discount changes the economics of a rich fixed preamble. Keep expectations straight on images: it reads photos and answers in words; it will not draw a revised part. Also note that Japanese-language quality and knowledge of domestically distributed materials are not covered by any of these indexes — a high composite score does not guarantee it knows a local vendor’s filament by product code. Before adopting, run your three most frequent real prompts head-to-head against your current model; that beats staring at rankings. The broader workflow map is in our AI 3D design roadmap.

What a free frontier model changes

The facts in one paragraph: 2.8 trillion total, 104 billion active — light to compute, heavy to hold. Composite score 57 at release, 60 in early August, fourth overall; first on AutomationBench-AA. Three dollars in, fifteen out, thirty cents cached. An open-weight — not open-source — license with two revenue triggers most users will never hit. And a 594-gigabyte floor that keeps it off your desk. What the release really did was raise the burden of proof for paid models: when something this capable is free to obtain and cheap to rent, charging flagship prices requires demonstrating flagship value. That pressure works in your favor. The one thing to do today costs five minutes: take three questions you actually ask, send them to Kimi K3 and to your current model, and compare. Your own tasks are the benchmark that matters.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.