Morning Edition · Monday, July 27, 2026Published at 1:32 AM EDT · New York

Moonshot AI Releases Kimi K3 Weights, a 2.8-Trillion-Parameter Open Model

The Chinese lab is publishing downloadable weights for a mixture-of-experts model that it says outperforms proprietary US systems on several coding and agent benchmarks. The company reported those benchmark figures itself.

Save

Moonshot AI Releases Kimi K3 Weights, a 2.8-Trillion-Parameter Open Model

Moonshot AI is releasing the full weights of Kimi K3 today, following the model's announcement on July 17. The company posted a countdown on its Hugging Face account ahead of the release. The architecture is a sparse mixture-of-experts model with 2.8 trillion total parameters, 16 of 896 experts active per token, a context window of one million tokens, and native multimodal input.

On Moonshot's own charts, K3 leads every tested system on Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, and Automation Bench. It trails what the company labels Fable 5 and GPT-5.6 on FrontierSWE and DeepSWE. Moonshot says the model beats Claude Opus 4.8 and GPT-5.5 on most of its internal tests. These are vendor-reported figures on a mix of proprietary and public benchmarks, and independent reproduction on standardized evaluations has not yet been published.

The significance is less any single score than the pattern. As analyst Nathan Lambert described it, this is an open-weights escalation: a Chinese lab shipping frontier-scale weights that anyone can download, fine-tune, and serve, at a moment when Washington restricts foreign access to closed US models. Running a 2.8-trillion-parameter model is not cheap, but making the weights public moves control away from a company's access controls and toward whoever owns the computing hardware.

Veracity: Corroborated

83/100

If true, who benefits: Moonshot AI and China's sovereign-compute push gain a downloadable frontier substitute that undercuts US closed-lab distribution, while GPU cloud and inference providers gain serving demand and the closed labs that sell access are pressured.
The nuance: The release, size, and open weights are independently confirmed by Tom's Hardware, but the parity claim rests on vendor runs that mix at least three agent harnesses, and independent reruns already come in lower (roughly 85% versus the reported 88.3% on Terminal-Bench 2.1).

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

Publishing weights at this scale gives governments and enterprises cut off from US application programming interfaces a downloadable frontier substitute they can host themselves. That erodes the closed labs' distribution advantage through deployment rather than through raw capability. The parties exposed are the closed labs that make money from access, meaning OpenAI and Anthropic, along with the specialized GPU cloud providers and sovereign-compute programs that benefit if inference for large open models has to run somewhere. The weakest point is the self-reported benchmarks. The claim of parity is only as strong as independent reproduction proves it to be.

What to watch

Independent SWE-bench Verified and LiveCodeBench runs on the released weights, which will show whether K3 holds parity outside Moonshot's own testing or whether the lead is an artifact of the benchmark.
Which cloud and inference providers set up hosted K3 endpoints, and at what token price. That will signal how much real demand exists for serving a 2.8-trillion-parameter open model rather than renting a closed API.

Observations to monitor, not financial advice.

3 sources

Synthesized from: Polylog editors · VentureBeat · Interconnects (Nathan Lambert)

Part of a tracked trend

Chinese Open-Weight Models Emerge as the Non-US AI Stack

As Washington restricts foreign access to US frontier models, governments and enterprises cut off from American AI increasingly standardize on downloadable Chinese open-weight models, splitting the world into competing AI supply blocs rather than a single frontier.