Kimi K3: Moonshot's largest open-model claim, graded on its own exam

Moonshot AI calls Kimi K3 the largest open-weight model ever built, and to make the case the company published a wall of benchmarks showing K3 ahead of several named rivals. Read two lines down in Moonshot's own announcement and the concession is right there: K3's "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol"^[1]. So the pitch is not that K3 is the best model on the market. It is Moonshot's claim to the best open model, pending a release that will let outside researchers inspect the artifact itself.

The catch is how it was measured: Moonshot selected the headline comparisons and reported many results through a mixture of its own and model-specific harnesses. The weights are not public yet. K3 is available through Moonshot's hosted products and API, but it cannot be independently run or audited locally. This is the read before the more complete review that July 27 may make possible.

What did Moonshot ship, and what did it hold back?

K3 is a 2.8-trillion-parameter mixture-of-experts model that routes each token through 16 of 896 experts, with a 1-million-token context window and native vision input^[1]. Its API uses reasoning and, at launch, defaults to max thinking effort; Moonshot says low- and high-effort modes will follow in subsequent updates^[1]. Moonshot prices it at $3 per million uncached input tokens and $15 per million output tokens, while cached input costs $0.30 per million tokens^[3].

Moonshot says the full weights and a technical report with further detail on K3's architecture, training, and evaluations will arrive by July 27, 2026^[1]. Until then, outside evaluators can inspect only the hosted model's behavior, not the artifact or training account behind it. That distinction matters: an API benchmark can be independently gathered, but it cannot establish what a released checkpoint will do under another team's inference stack, quantization, or agent scaffold.

A download will not make K3 locally practical for most teams. Moonshot recommends deploying it on supernode configurations with 64 or more accelerators^[1]. The release should let well-equipped labs and inference providers test the weights under different setups. It will not turn a 2.8-trillion-parameter model into a laptop project.

Can you trust the benchmarks?

K3's headline scores should be read as vendor-reported results, not a neutral ranking. Moonshot's table mixes its KimiCode harness with Claude Code, Codex, and results drawn from external leaderboards; on several tests, it reports the best available rival score across harnesses^[1]. Those choices do not make the results useless. They do make direct rank-order comparisons fragile, because agent scaffolding, token budgets, and tool configuration can move an outcome as much as the underlying model. We keep a .

The independent reads that exist are necessarily API-based: with no public weights, no one can run K3 locally in a neutral harness. Within that limit, the picture is strong. Artificial Analysis currently places K3 fourth of 186 models on its Intelligence Index, with a score of 57^[4]. Arena.ai's Frontend Code Arena put it first, at 1,679 points, ahead of Claude Fable 5^[5]. That is K3's clearest outside result, but it is one specialized arena, not a general verdict on coding.

There are caveats. Artificial Analysis calls K3 notably slow and very verbose, measuring 35.2 output tokens per second^[4]. Developer Simon Willison recorded a simple SVG task that used 13,241 reasoning tokens and cost about 25 cents^[6]. Moonshot's own architecture claims aim to improve efficiency at scale; the hosted behavior available today is what buyers have to judge.

A separate government evaluation offers a useful precedent, not a rule. When the US government's AI standards center tested DeepSeek V4 Pro, it recorded 74 percent on SWE-bench Verified and concluded:

"DeepSeek V4 scores better on DeepSeek's self-reported evaluations than on CAISI evaluations"^[7].

Vendor-reported results and outside evaluations can diverge, even for the same model. K3's weights have not yet been released, so outside teams cannot test that artifact themselves.

Is it even the cheap option?

Ask where Kimi K3 fits and the word that comes back is cheap. The price tag only half-agrees.

Against the US frontier, K3's $3-in, $15-out list price sits at the workhorse tier, not the flagship one. It undercuts Opus 4.8 at $5/$25 and Fable 5 at $10/$50. Its list price matches Claude Sonnet 5's $3/$15, but through August 31, 2026, Sonnet's introductory rate is $2/$10^[8]. Today, then, K3 costs 50 percent more than Sonnet on both input and output. The trade on offer is not frontier capability at a discount. It is a model Moonshot says trails the two leading proprietary systems, priced above a temporarily discounted mid-tier alternative.

DeepSeek V4 Pro makes the sharper open-model comparison. On Artificial Analysis's same Intelligence Index, K3 scores 57 while DeepSeek V4 Pro at max effort scores 44^[4][12]. That is a meaningful capability gap in the available API evaluation, not a footnote to be hidden. OpenRouter lists DeepSeek at $0.435 per million input tokens and $0.87 per million output tokens, roughly seven times cheaper for input and seventeen times cheaper for output than K3^[9].

DeepSeek's weights are publicly downloadable under an MIT license, and CAISI has evaluated the released model^[10]. Artificial Analysis lists it as a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters per token^[12]. That does not settle a capability contest. It makes the tradeoff plain: K3 looks stronger in the available API tests; DeepSeek is cheaper, faster in Artificial Analysis's measurements, downloadable, and already exposed to outside testing.

K3 may prove more capable when its weights arrive. For now, the available API evaluations make it a serious contender, but they do not erase the capability-price-reproducibility tradeoff.