Kimi K3 is a 2.8 trillion-parameter AI model from Moonshot AI, and the full weights are due in the open by 27 July 2026.
What Moonshot AI actually shipped
Moonshot AI released Kimi K3 on 16 July 2026 with a claim worth checking: the world's first open-source model in the three-trillion-parameter class. The Kimi K3 documentation gives the specification. Native visual understanding. A one-million-token context window. Thinking always on, with reasoning effort dialled to low, high, or max, and max as the default.
The commercial terms arrived first, the weights follow. Kimi K3 costs $3.00 per million input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens. Moonshot AI gates flagship access behind a successful top-up of at least one dollar. Full model weights land by 27 July 2026, with a technical report covering architecture, training, and evaluation to follow.
Moonshot AI also makes a claim about consistency rather than a single launch. Across nine of the twelve months from July 2025 to July 2026, Kimi models held the frontier in open-source model scale. Sustained leadership across a year reads differently from a one-off record.
The architecture is the real story
Being the biggest open model is a weak signal on its own. Kimi K3 matters because Moonshot AI made 2.8 trillion parameters affordable through three changes working together, and the company published two of the underlying techniques as papers with code.
The first change is Kimi Delta Attention, a hybrid linear attention mechanism. Standard attention compares every new token against every previous token, and cost grows accordingly across a long input. Kimi Delta Attention instead maintains a fixed-size memory, updated as the model reads, deciding at each step which older information to overwrite. Across a million-token input, the difference between quadratic comparison and fixed-size memory is the difference between a demo and a product.
The second change is Attention Residuals. In a standard transformer, each layer's output is added to a running total with equal weight, so an individual layer's contribution dilutes as depth increases. Attention Residuals apply attention across depth itself, letting each layer draw selectively on the outputs of earlier layers — much as ordinary attention draws selectively on earlier tokens. Moonshot AI groups layers into blocks and attends over those block outputs.
The third change is sparsity. Under a framework Moonshot AI calls Stable LatentMoE, Kimi K3 activates just 16 of 896 experts per token. Fewer than two per cent of the available experts fire on any given token.
Roughly 2.5 times the overall scaling efficiency of K2, converting compute into capability more effectively.
— — Moonshot AI, on Kimi K3
Read that number as a statement about constraints. Moonshot AI operates under compute limits that American laboratories do not face, and the response was architectural creativity rather than a larger cluster. Research published openly, with code, means every developer inherits the gain — including the laboratories with no shortage of chips.
Where Kimi K3 lands against the frontier
Independent measurement puts Kimi K3 third on Artificial Analysis' Intelligence Index at 57 points, and first among open models by a clear margin — the closest open-weight competitor, Z.ai's GLM-5.2, scores 51. On Arena.ai's Code Arena WebDev leaderboard, where developers pick the better of two outputs for the same frontend task, Kimi K3 debuted in first place at 1,679 Elo.
Cost per task is where the pressure shows. Kimi K3 completed the Intelligence Index at $0.95 per task, against $1.04 for GPT-5.6 Sol and $2.75 for Claude Fable 5. Analysis of those three numbers explains the anxiety in San Francisco better than any benchmark chart: comparable output, at roughly a third of the price, with weights you can host yourself.
Moonshot AI is also honest about the gap. The company concedes Kimi K3 still trails Claude Fable 5 and GPT-5.6 Sol in head-to-head performance. Demonstrations lean on long-horizon agentic work — autonomous GPU compiler development, a chip design completed in 48 hours, automated video editing — rather than on claims of outright superiority. Infrastructure has moved in step: vLLM shipped production-scale Kimi K3 support in preview on 22 July 2026.
Why open weights change the calculation
Whoever holds the weights sets the usage policy. Evidence for how much that matters arrived within days of launch, when Hugging Face disclosed that commercial model guardrails had blocked forensic analysis of a live intrusion, and the defence ran on an open-weight model hosted in-house — a sequence I unpack in the Hugging Face breach and the guardrails that blocked the defence.
For anyone building outside the United States and China, open weights are the difference between renting capability and holding it. A model running on your own infrastructure keeps sensitive data inside your jurisdiction, survives an export-control decision made in another capital, and can be fine-tuned on languages and tasks no frontier laboratory has commercial reason to prioritise. Emergent Intelligence (EI) — the dignity-first frame I use for what most people call AI — treats that capacity as a precondition for agency rather than a technical nicety.
One honest caveat belongs here. Moonshot AI has not disclosed the training datasets, the training methods, or the licence attached to the promised weights. Open weights are not the same as open data or open process, and the gap deserves naming rather than applause. Until the licence appears, Kimi K3 is a model you can download, not yet a model you can fully audit.
Frequently Asked Questions
These are the questions people are asking about Kimi K3 and open AI models at frontier scale. Short answers follow, drawn from Moonshot AI documentation and independent benchmarks.
What is Kimi K3?
In short, Kimi K3 is a 2.8 trillion-parameter multimodal AI model released by Moonshot AI on 16 July 2026, with native visual understanding and a one-million-token context window. Moonshot AI documentation describes Kimi K3 as the first open-source model in the three-trillion-parameter class, with full weights due by 27 July 2026.
How does Kimi K3 achieve its efficiency?
According to Moonshot AI, three changes combine: Kimi Delta Attention, a hybrid linear attention mechanism using fixed-size memory rather than comparing every token pair; Attention Residuals, which let each layer draw selectively on earlier layers; and heavy sparsity, activating 16 of 896 experts per token. Company data puts the combined result at roughly 2.5 times the scaling efficiency of Kimi K2.
Why is an open model at this scale significant?
The key is control. Analysis from Artificial Analysis ranks Kimi K3 third overall at 57 points and first among open models, at $0.95 per task against $2.75 for Claude Fable 5. Comparable capability at a third of the cost, with downloadable weights, removes most of the reasons a developer would default to a proprietary frontier model.
Who benefits most from open AI weights?
In other words, everyone who cannot rely on someone else's terms of service. Evidence from the July 2026 Hugging Face incident shows defenders locked out of hosted models mid-breach, while organisations outside the United States and China gain data residency, resilience against export controls, and the ability to fine-tune for languages no frontier laboratory prioritises.
What are the limits of the Kimi K3 release?
The answer is transparency. Moonshot AI has not disclosed training datasets, training methods, or the licence governing the promised weights, and the company concedes Kimi K3 still trails Claude Fable 5 and GPT-5.6 Sol head to head. Simply put, open weights are not open data, and research reproducibility remains incomplete until the technical report lands.
