Kimi K3 Is Third Best at Coding. It Is Also Not Open.

Kimi K3 scores 93.4% on SWE-bench Verified, the third best result among the 72 systems on vals.ai's independent leaderboard, and close enough to Claude Fable 5 that the two are a statistical tie. Moonshot bills K3 as the largest open model ever built. As of today the weights are not downloadable anywhere, and the one independent evaluator that classifies models lists it as proprietary.

• 93.4% ±1.11 on an independent run. Only GPT-5.6 Sol (96.2%) and Claude Fable 5 (95.0%) scored higher, out of 72 systems tested on the same harness.
• The gap to Fable 5 is not real. 1.6 points against a combined error margin of 1.5 is roughly one sigma. K3 lists at $3 per million input tokens; Fable 5 lists at $10.
• Nothing is open yet. Hugging Face's moonshotai organisation holds 18 repositories and none is K3, github.com/MoonshotAI/Kimi-K3 returns a 404, and Moonshot's blog promises weights "by July 27, 2026" with no licence stated.
• The vendor's own number was high. Moonshot claimed 88.3% on Terminal-Bench 2.1. Artificial Analysis ran the same benchmark itself and measured 85.02%.

What did the independent test actually measure?

vals.ai runs every model itself through one minimal harness called mini-swe-agent, on the 500 human-validated GitHub issues in the SWE-bench Verified split. The model gets a single tool, bash, and has to navigate the repository, find the bug and produce a patch that passes the project's own unit tests. Every system on the board gets the identical setup.

RelatedQwen3.7-Max Tops SWE-Bench Pro and Terminal-Bench 2.0

That detail carries most of the weight here. SWE-bench scores a model and its scaffolding together, which is why a lab can raise its published number by improving the agent wrapped around the model rather than the model itself. A neutral harness strips that variable out.

K3 came in at 93.40% with a reported error margin of ±1.11, ranked third of 72. Read the ordering carefully, because two of the three comparisons around it are noise. The 0.4 points over GPT-5.6 Luna sit inside both error bars and mean nothing. The 1.6 points behind Claude Fable 5 work out to about one sigma, which is also not a real separation. The result that does hold up is the 4.8 points over Claude Opus 4.8, at roughly 2.7 sigma. In plain terms: K3 belongs in the same conversation as the frontier models from Anthropic and OpenAI, and it clearly beats Opus 4.8 on this test.

Why is it wrong to call Kimi K3 open?

Moonshot's own announcement calls K3 "the world's first open 3T-class model" and commits to releasing "the full model weights by July 27, 2026." That is a promise with a deadline, not a shipped artifact, and the distinction matters to anyone planning around it.

Checked on July 19: the moonshotai organisation on Hugging Face lists 18 repositories and not one of them is K3, with the newest official model artifact being Kimi-K2.7-Code from June 15. The URL github.com/MoonshotAI/Kimi-K3 returns a 404. Moonshot's own GitHub organisation still describes Kimi K2.5 as its most powerful open-source model. Artificial Analysis, which runs its own evaluations, answers the question directly in its model FAQ: Kimi K3 is proprietary, and the weights are not publicly available. It excludes K3 from its open-weights listing while including Kimi K2.6.

No licence has been stated either. Moonshot has not said whether the release will be full weights or something narrower, such as a base-only or gated tier. A third-party site claims a modified MIT licence is coming, but nothing on any Moonshot channel confirms that, so treat it as rumour.

A correction to our own coverage. Our leaderboard had K3 flagged as open weights until this morning, which pushed it to the top of our own "best open-source AI coding model" page as a model nobody can download. That is fixed: K3 now sits on the board as a closed model, and that page correctly leads with DeepSeek V4 Pro. Our July 17 story also called K3 "the largest open model ever," which carried the same error, and we are correcting it here rather than quietly. If you want the strongest model you can actually run on your own hardware today, GLM-5.2 leads Artificial Analysis's open-weights index at first of 97 under a genuine MIT licence, and DeepSeek's V4-Pro weights have already been downloaded about 1.49 million times.

What did Moonshot overstate?

At launch Moonshot led with Terminal-Bench 2.1 at 88.3%, measured on its own KimiCode harness. Artificial Analysis ran Terminal-Bench 2.1 against K3 itself and got 85.02%. The vendor's number was 3.3 points high.

That is not an accusation of bad faith, and it is not unusual. It is the same pattern we track across our whole leaderboard, where vendor-published scores run between 2.6 and 11.6 points above what a neutral harness produces for the same model. It is a good reason to wait for somebody else to run the benchmark before treating a launch slide as a measurement.

The independent placements are also more modest than the launch framing suggests. Artificial Analysis has K3 fourth on GDPval-AA v2 with an Elo of 1684, and fourth of 187 models on its Intelligence Index v4.1 at 57.1. It does rank second on AA-Briefcase, an agentic benchmark, at an Elo of 1545. Fourth and second are strong results for a model this new. They are not the top of the table.

Why do developers rank it first when benchmarks rank it third?

Here is the genuinely interesting split. On Arena's WebDev leaderboard, where developers compare two anonymous model outputs side by side and vote, Kimi K3 ranks first of 99 models with a score of 1679 across 1,757 votes, ahead of every model that beats it on SWE-bench.

Both numbers are real, and they measure different things. SWE-bench asks whether a generated patch makes a failing unit test pass, which is a question about correctness on existing codebases. Arena asks which of two outputs a developer would rather have, which folds in code style, structure, explanation and how the result looks when it renders. A model can be excellent at producing likeable, well-organised front-end code and merely very good at surgically fixing a bug in someone else's repository.

RelatedClaude Sonnet 5: Near-Opus Coding at Half the Price

The practical read: if your work is closing tickets in a large existing codebase, weight the SWE-bench figure. If it is generating new interfaces and pages from scratch, the Arena result is closer to your job.

What does it actually cost?

Moonshot lists K3 at $3.00 per million input tokens and $15.00 per million output, with cached input at $0.30, confirmed on Moonshot's own pricing page on July 19. That is roughly triple Kimi K2.6 at $0.95 and $4.00. A launch top-up rebate running to August 12 pays vouchers worth 10% to 30% of a top-up, which lowers your effective bill but does not change the per-token rate.

Two caveats before anyone builds a cost model. vals.ai publishes a cost-per-test figure for most systems on its board and lists K3's as not available, so a clean like-for-like cost per solved task is not yet possible, and any figure you see quoted as one has been reconstructed rather than measured. Second, K3's reasoning_effort parameter is currently locked to "max", so the output tokens, the expensive half of the bill at $15 per million, are not under your control the way they are on models that let you dial reasoning down.

Our take

The score is real and the model is genuinely good. A system that ties Claude Fable 5 inside the error bars while listing at a third of the input price is a serious piece of engineering, and the Arena result suggests developers like working with it more than the benchmark ranking implies.

Two things still need saying. The word "open" is doing heavy lifting in the marketing and none in practice, and we amplified that ourselves until this morning, which is why the correction above is in the piece rather than buried in a changelog. And when a lab's own headline benchmark lands 3.3 points high under a neutral harness, the reasonable response is to wait for independent numbers on the rest of the claims too.

Use the API if the price fits your workload. Do not build a self-hosting roadmap around a repository that currently returns a 404. Ask again on July 28.

Original analysis by GenZTech. Scores quoted from independent evaluations run by vals.ai and Artificial Analysis; pricing confirmed against Moonshot's own documentation on July 19, 2026. Corrects our July 17 report, which described Kimi K3 as an open model before confirming that no weights had been published.