>_TheQuery
// Reading nowStart
← All Articles

Kimi K3 Is the Largest Open-Weight Model Ever Announced. It Also Beats Claude Opus 4.8 and GPT-5.5.

By Addy · July 17, 2026

Moonshot AI released Kimi K3 on July 16 through its API, Kimi apps, and Kimi Code. The full open weights -- the downloadable files that would let anyone inspect, modify, or self-host the model -- are not out yet. Moonshot says they will arrive by July 27.

That distinction matters. K3 is available as a product today, but its status as the largest open-weight model ever released is still a promise. The architecture cannot be audited, the launch numbers cannot be reproduced from the weights, and the technical report is due alongside them. This publication has documented what happens when a vendor-run benchmark meets a harder independent harness. Every number below should be read with that caveat attached.

The caveat does not make the launch unimportant.

At 2.8 trillion total parameters, K3 is the largest model ever announced for an open-weight release -- roughly 75% larger than DeepSeek V4 Pro's 1.6 trillion parameters. Independent API evaluation by Artificial Analysis gives K3 a score of 57 on its Intelligence Index and places it third overall, behind Claude Fable 5 and GPT-5.6 Sol, but in the same tier as Claude Opus 4.8 and GPT-5.5. Blind testing by Arena also put K3 first for front-end coding, ahead of every leading American model tested.

That is the honest headline: K3 does not establish overall dominance, but it does beat proprietary flagships on important coding evaluations before its weights are even downloadable.

What Actually Shipped

K3 is a sparse mixture-of-experts model with 896 total experts and 16 activated for each token. Moonshot pairs that scale with Kimi Delta Attention and Attention Residuals, two mechanisms designed to move information efficiently across both sequence length and model depth. Its Stable LatentMoE framework addresses a basic problem created by extreme sparsity: keeping experts balanced instead of overloading a few while leaving others underused.

The model has a 1 million token context window and native vision. Moonshot reports a score of 90.4 on BrowseComp when K3 uses the full context without compaction or other context-management techniques. That result is vendor-run, not independent proof that every task remains reliable at one million tokens, but it is more useful than simply printing "1M context" on a specification sheet.

Reasoning is always on at launch. K3 currently runs at maximum inference-time compute, with lower-effort modes promised later. That matters for both latency and cost: a cheap per-token rate does not guarantee a cheap task if the model generates a long reasoning trace.

Moonshot also used quantization-aware training from the supervised fine-tuning stage onward, with MXFP4 weights and MXFP8 activations. The goal is broad hardware compatibility once the weights arrive. Broad does not mean small. Moonshot recommends a deployment with at least 64 accelerators, making K3 open-weight in the inspect-and-control sense, not in the runs-on-your-laptop sense.

The Benchmarks, With the Harnesses Left In

The cleanest result is the independent one. Artificial Analysis tested K3 through Moonshot's API and scored it 57 on its composite Intelligence Index. That makes K3 the third-ranked model in the current table and, if the weights ship as promised, the clear leader among open-weight models. It also shows the limit of the launch headline: K3 sits alongside Opus 4.8 and GPT-5.5 overall rather than simply beating both across every category.

The coding results are stronger.

On DeepSWE, Moonshot reports 67.5% using Kimi Code. The official DeepSWE setup scores K3 at 67.3% with the benchmark's mini-SWE-agent harness. A two-tenths gap is reassuring because it suggests the result is not being created by Moonshot's own agentic wrapper. It also places K3 ahead of Opus 4.8 on the evaluation this publication currently treats as more rigorous than SWE-bench Pro, though behind GPT-5.5 and GPT-5.6 Sol.

On Terminal-Bench 2.1, K3 scores 88.3, narrowly behind GPT-5.6 Sol at 88.8 and ahead of the 84.6 Moonshot cites for both Claude Fable 5 and Opus 4.8. The comparison is useful but not clean. K3 ran through Kimi Code, the Claude models through Terminus 2, and the GPT models through Codex. A benchmark score measures the model, the harness, and their interaction. Pretending those three components are interchangeable would make the table easier to market and less useful to developers.

The most independent coding win came from Arena rather than Moonshot. In blind preference tests, K3 reached number one on Frontend Code Arena, placing first across six of seven front-end domains and jumping 17 positions from Kimi K2.6. That result matters because users judged the outputs without being told which model produced them. K3 did not merely clear Opus 4.8 and GPT-5.5 there. It also beat Fable 5 and GPT-5.6 Sol on that specific kind of work.

The model still loses on important measures. Fable 5 leads K3 on long-horizon knowledge work in Artificial Analysis, while GPT-5.6 Sol stays ahead on the overall Intelligence Index and several deeper reasoning tests. The pattern is not "K3 beats everything." It is more specific: K3 is already in the proprietary frontier tier overall and can lead it on coding tasks.

The Demonstrations That Matter More Than a Leaderboard

Moonshot's most interesting evidence is a 24-hour GPU-kernel optimization exercise. Each model received four tasks in an identical sandbox and could profile, rewrite, and benchmark its work with minimal supervision. K3 performed competitively with Fable 5 and ahead of Opus 4.8, GPT-5.6 Sol, and GPT-5.5 in Moonshot's run.

K3 also built MiniTriton, a compact GPU compiler with its own intermediate representation, optimization passes, and code-generation pipeline. The result reportedly supported an end-to-end nanoGPT training run, not just isolated toy kernels.

These remain Moonshot case studies. They are not independent evaluations, and the company has every incentive to select tasks that make its model look strong. But they test something most leaderboards do not: whether an agent can sustain coherent engineering work across hours, tools, failed attempts, and verification steps. That is closer to the capability developers are actually buying than a single-response coding puzzle.

The Price Advantage Is Real. It Is Also Smaller Than It Looks.

K3 costs USD 0.30 per million cached input tokens, USD 3 per million cache-miss input tokens, and USD 15 per million output tokens through Moonshot's API. Moonshot says its Mooncake serving system reaches cache-hit rates above 90% on coding workloads, which would make repeated work over the same repository substantially cheaper.

The sticker prices undercut Opus 4.8, especially on cached input. They do not translate into a five-to-eight-times advantage on a complete task.

Artificial Analysis measured K3 at USD 0.94 per Intelligence Index task. Opus 4.8 cost USD 1.80, so K3 was roughly half the price. GPT-5.6 Sol cost USD 1.04, making K3 only slightly cheaper. K3 is also expensive by open-weight standards: GLM-5.2 cost USD 0.32 per task in the same analysis and DeepSeek V4 Pro cost USD 0.04.

The reason is output volume. K3 generated roughly 132 million output tokens across the evaluation suite. That was more efficient than Kimi K2.6, but still verbose enough to eat into the lower token price. Always-on maximum reasoning can do the same thing in production.

So the pricing question is not why American labs charge five to eight times more for the same capability. The evidence does not support that framing. The better question is why Opus 4.8 costs roughly twice as much per measured task when K3 can match its overall intelligence score and beat it on several coding evaluations. That gap is smaller, but it is real.

The Open-Weight Claim Is Now the Test

K3 already has more independent evidence behind it than the usual launch announcement. Artificial Analysis ran it through a broad evaluation suite. Arena users preferred its front-end work in blind tests. The DeepSWE score barely moved when the harness changed.

What nobody outside Moonshot can test yet is the part that makes K3 structurally different from Opus or GPT: the weights themselves.

If Moonshot releases the full 2.8-trillion-parameter checkpoint by July 27, under a license that permits real inspection, fine-tuning, and deployment, K3 will become the strongest open-weight model Artificial Analysis has measured and the largest one anyone can download. Researchers will be able to audit Kimi Delta Attention, test the Stable LatentMoE claims, reproduce the launch numbers, and find out whether the model behaves the same outside Moonshot's serving stack.

If the date slips, the license restricts meaningful use, or independent deployment produces materially weaker results, the phrase "open-weight" will have done ten days of marketing work before becoming a verifiable fact.

That is why the provisional verdict matters. Kimi K3 is already a frontier model. It is already a serious coding model. It is already cheaper than Opus 4.8 in independent evaluation, though not dramatically cheaper than GPT-5.6 Sol.

The remaining claim is the largest one: that a 2.8-trillion-parameter model this capable is about to become something anyone can inspect and control rather than merely rent through an API.

The weights are due July 27. That is when the most important Kimi K3 benchmark begins.

Sources:

Previously on TheQuery: