Kimi K3: The Largest Open Model Yet, and the End of Rock-Bottom Chinese AI Pricing

On July 16, 2026, Moonshot AI released Kimi K3, a 2.8 trillion parameter model it describes as the first open model in the 3-trillion-parameter class, with full weights due by July 27. Two things make it worth a careful read rather than a quick headline. First, independent testing places it at the frontier tier, close to the strongest closed US models, which is unusual for any open-weights release. Second, and more revealing, it is priced like a Western mid-tier model rather than a bargain, a break from the rock-bottom pricing that Chinese labs used to win attention. This piece walks through what K3 is, how it actually performs by independent measurement, and why its price tag may be the most important thing about it.

Moonshot AI

What Kimi K3 is, technically

Kimi K3 is a very large sparse mixture-of-experts model. Its 2.8 trillion total parameters set the ceiling on what it can know, while only about 16 of its 896 experts fire on any given token, which keeps the compute cost per token far below what the raw parameter count implies. That total-versus-active split is the standard recipe for modern efficient models, taken here to an unusually large scale.

The architecture is where Moonshot did new work rather than just scaling up. K3 introduces Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), both aimed at helping information flow cleanly through very long sequences and deep networks, along with a Gated MLA design and a Stable LatentMoE framework for routing. It uses MXFP4 weights with MXFP8 activations, a low-precision format that makes a model this size more practical to serve. The headline capability that falls out of all this is a 1 million token native context window, roughly four times the 256K of the previous Kimi K2.6, enough to hold an entire codebase or a long multi-step agent run in a single context.

How it performs, by independent measurement

Moonshot published strong self-reported numbers, including 88.3 on Terminal-Bench 2.1 and 93.5 on GPQA Diamond. As always, the launch chart flatters the lab that publishes it, so the more useful reference is independent evaluation.

On the Artificial Analysis Intelligence Index, an independent composite, Kimi K3 scores 57 and ranks fourth of 189 tracked models. That places it directly between two closed US frontier models, above Claude Opus 4.8 at 56 and below GPT-5.6 Sol at 59, with Claude Fable 5 on top at 60. The significance is not that K3 wins; it does not. It is that an open-weights model a developer can download is now sitting inside the same three-point band as the leading closed models, rather than a tier below them. On agentic knowledge-work benchmarks the picture is similar: K3 lands just behind the top closed models and ahead of some of them, depending on the test.

Two caveats keep this honest. Independent evaluators also found K3 unusually verbose, generating around 130 million output tokens across the Intelligence Index run against a median of 63 million, and relatively slow at about 62 tokens per second. Verbosity and latency both cost money and time in production, and a composite score does not capture them. A model that reaches a frontier-level answer but takes more tokens and more seconds to get there is not straightforwardly interchangeable with one that does not.

The price is the real story

For two years, the reliable pattern with Chinese models was simple: competitive capability at a fraction of the price. Kimi K3 breaks that pattern deliberately.

K3 lists at $3.00 per million input tokens and $15.00 per million output tokens, with a $0.30 cache-hit input rate. Its predecessor, Kimi K2.6, listed at $0.95 input and $4.00 output. That is roughly a tripling of price in a single generation, and it lands K3 exactly on top of a Western mid-tier model: Claude Sonnet 4.5 also lists at $3.00 and $15.00. Measured per task, K3 runs about $0.94, against roughly $0.04 for DeepSeek V4 Pro and $0.32 for Z.ai's GLM-5.2, so it is now many times more expensive than the cheapest capable open models, while still undercutting the top Western systems.

The strategic read is that Moonshot is no longer competing on being the cheapest. It is competing on being good, and pricing accordingly. That is a meaningful shift in posture. When a frontier-capable Chinese open model prices itself at parity with a Western mid-tier model, it is signalling confidence that buyers will pay for capability rather than defaulting to whatever is cheapest, and it suggests the era of Chinese frontier models as automatic bargains is closing. Whether the market rewards that confidence is the open question; there is a real risk that price-sensitive users simply route to the cheaper open alternatives that still exist.

Open weights at this scale changes the calculus

The other half of the story is that K3 is not just capable and fairly priced; it is open. Full weights are due July 27, which means an organisation can in principle download a frontier-tier model and run, inspect, fine-tune and own it rather than renting it through an API. At 2.8 trillion parameters the practical barrier is real, since serving a model this large takes serious infrastructure, but the option exists, and it did not exist at this capability level for an open model before.

This continues a trend rather than starting one. The open-weights field has been closing on the closed frontier all year, from strong Chinese releases to US efforts like Thinking Machines' Inkling. K3 is the most striking data point yet because it is the largest and among the most capable, and because it declines to compete on price. Taken together, these releases make hard-wiring a product to a single closed model look increasingly like a choice rather than a necessity.

What it means for people building on models

For a team choosing what to build on, Kimi K3 widens an already wide menu. It adds a frontier-tier option that is open, competitively but not cheaply priced, and strong at long-horizon coding and agentic work. It is not the obvious pick for latency-sensitive or extremely cost-sensitive workloads, given its verbosity and speed, and it is not the cheapest way to get open weights. It is a strong all-rounder that happens to be downloadable.

The harder question is the recurring one: with the frontier now crowded by closed US models, open US models, and open Chinese models that trade blows on capability and diverge sharply on price, how do you exploit that without rebuilding every quarter? The durable answer is to stay portable and route each job to whatever fits it, rather than committing a whole stack to one provider whose price, availability or standing can change. A model-agnostic workspace such as Metir AI, which already offers Kimi K3 alongside closed frontier models, is one way to hold that flexibility, so trying K3 on an agentic task or a cheaper model on a bulk one is a routing choice rather than a migration.

The bigger picture

Kimi K3 is two stories in one model. The first is a straightforward capability milestone: the largest open-weights model yet, independently measured into the frontier tier, and further evidence that the gap between open and closed is narrow and shrinking. The second is quieter and arguably more consequential: a leading Chinese lab has chosen to price a frontier model like a Western one, betting that the market now values its capability enough to pay for it.

If that bet pays off, it marks a genuine repositioning of Chinese AI from price challenger to capability peer. If it does not, cheaper open models are one API call away. Either way, the direction of travel is clear: the frontier is filling with credible, downloadable options, and the advantage increasingly goes to whoever can use the best one for each job without being locked to any of them.

Sources:

• Kimi K3: Open Frontier Intelligence | Moonshot AI
• Kimi K3 Intelligence, Performance and Price Analysis | Artificial Analysis
• China's Moonshot AI releases Kimi K3, the largest open-source model ever | VentureBeat
• Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI | The Decoder
• Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model | MarkTechPost