# Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model That Tops Open Rankings

The Beijing lab's mixture-of-experts system trails only the closed frontier of Claude Fable 5 and GPT-5.6 Sol while beating Claude Opus 4.8 on coding and agent benchmarks, and its weights are a 1.4-terabyte free download.

- Published: 2026-07-28T05:31:39.988Z
- Canonical: https://polylog.news/ai/2026-07-28/moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open-w
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news)

Moonshot AI, the Beijing startup backed by Alibaba, has published the open weights for Kimi K3, a 2.8-trillion-parameter model it calls the largest ever released for download. The [weights went live on Hugging Face](https://t.me/ai_machinelearning_big_data/10601) on July 26, a day ahead of schedule. The design is a mixture-of-experts model with a context window of roughly one million tokens and an always-on reasoning mode. And, [as VentureBeat reported](https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems), only about 32 billion parameters activate for each token, which holds the cost of running the model far below what its full size suggests.

Moonshot's own evaluations place K3 behind the closed frontier of Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, but ahead of every other model it tested, including Claude Opus 4.8 and GPT-5.5, on coding and agentic suites. [Tom's Hardware reported](https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3) that K3 beat Claude Fable 5 on the Frontend Code Arena benchmark. A widely circulated [AI Post summary](https://t.me/aipost/7652) put K3 at the top of the open-weight tier of Artificial Analysis's Intelligence Index with a score of 57, a figure that reflects a single aggregator rather than independent reproduction of each named benchmark.

The practical constraint is deployment size. The full model is roughly 1.4 terabytes even when compressed to a lower-precision format (MXFP4 quantization), which makes self-hosting impractical for most teams and pushes real usage toward hosted endpoints and cloud providers. The distance between a model that can be downloaded and one that can actually be run is now the central economic question for open weights at this scale.

## What this means

The open-weight tier now includes a model that beats a two-month-old closed frontier model from a leading United States lab, which shortens the period during which a closed lab can charge a premium for capability alone. The exposed parties are sellers of metered application programming interfaces (APIs) whose pricing assumes a durable capability lead, and the beneficiaries are enterprises and government buyers who can now build a stack on downloadable Chinese weights. The 1.4-terabyte footprint means the near-term beneficiaries on the serving side are cloud hosts, not individual machines.

## What to watch

- Independent reproduction of the coding and agent numbers on Artificial Analysis, LMArena, and SWE-bench, since a single aggregator score is not the same as verified parity on each benchmark.
- How fast third-party inference providers create cheap hosted K3 endpoints, which determines whether the release is used in production or remains only a published specification.