FULL STORY

Kimi K3: From Anticipation to Topping Open-Source Charts

Moonshot AI's Kimi K3 launched its 2.8T parameter MoE model to immense anticipation, quickly topping the Hugging Face trending charts and sparking widespread ecosystem engagement.

2026-07-24 ~ 2026-07-28 · 5 episodes · 192 posts

Episode 1 · Moonshot's Kimi K3 Set for Release with Multi-Platform Support (2026-07-24, 4 posts)

Moonshot's new Kimi K3 model is expected to release next Monday with immediate availability on platforms like Together AI. Inference provider Baseten is also offering a $25 free API credit for developers.

Episode 2 · Moonshot AI Unveils Trillion-Parameter Kimi K3 (2026-07-24, 4 posts)

Moonshot AI released Kimi K3, a frontier open-source model with 2.8 trillion parameters utilizing sparse MoE and Delta Attention. The model supports one million token context multimodally, refuting clone model claims and highlighting massive engineering efforts.

Episode 3 · Kimi K3 Open Weights Launch Draws Instant Ecosystem Support (2026-07-26, 24 posts)

Moonshot AI took Kimi K3 through a tightly watched release cycle from teaser to open-weights launch. Based on the materials in this cluster, Kimi K3 was released on schedule around July 27, 2026 in open-weights form, with the release described as including model weights and a paper. The story matters not only because of the model itself, but because deployment platforms reacted immediately and developers began debating whether K3 could become a practical alternative in real workflows.

Confirmed

  • Moonshot’s official Kimi account first teased that Kimi K3 would be available as open weights soon. Hugging Face then showed an official-looking “Upcoming release” page for moonshotai/Kimi-K3, with a countdown targeting July 27, 2026 and a waitlist/notification entry point.
  • Multiple screenshots shared by community members showed substantial interest before launch, with the page displaying roughly 2,758 to 3,196 people waiting.
  • Community author nrehiew described the date as the release day for K3 weights and paper, and the existing cluster summary states that downloadable weights and the accompanying paper were made available after launch. yangyi later said Kimi K3 had indeed been open-sourced on schedule.
  • Modal announced Day 0 support, indicating that inference and deployment tooling would be available as the model went live.
  • Several posts noted that K3 had already been accessible via Kimi’s own product and API; the key change here was the release of downloadable weights into the open-source ecosystem.

Unconfirmed

  • The posts do not provide full details on licensing, parameter count, benchmark results, or exactly when every third-party platform would expose K3 broadly.
  • Existing materials mention community observations that the Hugging Face countdown page may have briefly disappeared, but the cluster does not contain enough evidence to determine whether that was a routine update or something more meaningful.

Why it matters

  • bindureddy argued that enterprises could consider shifting workloads previously handled by Sonnet and Terra to K3, and potentially fine-tuning it into custom frontier-grade systems.
  • HotExample4456 pointed out that local deployment remains difficult for many users, making inference-provider adoption a more practical near-term indicator of usefulness. SwordfishGreedy1945 framed the central test more sharply: whether developers will actually replace Claude or Codex with a Chinese model in real workflows.
  • The existing summary also captures a broader industry reading: theteknosaur saw K3 as another sign that the gap between top open and top closed models is shrinking, while markk treated it as part of a notably active week for AI releases.

4 more related posts →

Episode 4 · Moonshot releases Kimi K3 open weights amid license debate (2026-07-26, 155 posts)

Moonshot AI has published Kimi K3’s open weights and technical report on Hugging Face. Kimi K3 is a 2.8T-parameter MoE model with 104B active parameters, native vision support, and a 1M-token context window. It stands out as one of the largest open-weight releases to date, but the model’s deployment cost and its commercial licensing terms quickly became as notable as the model itself.

Confirmed

  • Moonshot describes Kimi K3 as a 2.8T MoE model with 896 experts, activating 16 experts per token for 104B active parameters.
  • The technical materials highlighted in posts say K3 introduces Kimi Delta Attention, Attention Residual, and a three-axis scaling design across sequence length, network depth, and model width. A lower-bounded decay floor of -5 was also added to chunkwise KDA to reduce finite-precision overflow issues.
  • Posts relaying the technical report say K3 scored 91.2 on BrowseComp. The same report was also described as finding 16 previously unknown vulnerabilities across six projects, including two Linux kernel vulnerabilities.
  • Community deployment estimates put the weights at about 1.4TB. One widely shared calculation said fully fitting the model would require at least 18 GPUs, or 8 B300 cards; Reddit users argued that even multi-node local workstations built around 2–4 RTX 6000 Blackwell systems would still struggle to run it natively.
  • The model uses a Modified MIT license. Posts summarizing the terms say local self-hosting is allowed, but if Kimi K3 is resold as a cloud or Model-as-a-Service offering by a company or affiliate with more than $20M in revenue over any rolling 12-month period, a separate agreement with Moonshot is required.

Unconfirmed

  • Beyond the reported benchmark and security results, Kimi K3’s real-world general-purpose performance still needs broader independent testing.

Why it matters

  • Kimi K3 pushes the scale of open-weight releases into the 3T class while combining long context, multimodality, and agent-oriented positioning, making it a notable signal in the frontier model race.
  • At the same time, the release exposes a familiar tension in open-weight AI: the model is publicly downloadable, but the hardware needed to run it is out of reach for most users, and the commercial terms sparked pushback. In particular, @Eyelbee argued that the licensing and pricing implications make third-party inference economics look too close to a closed API, weakening the practical openness of the release.

135 more related posts →

Episode 5 · Moonshot's Kimi K3 Tops Hugging Face Trending (2026-07-27, 5 posts)

On July 27, Moonshot's Kimi K3 model officially appeared on the Hugging Face platform, quickly drawing massive attention. The model garnered over 4,000 likes within approximately 30 minutes of its release, rapidly climbing to the top of the Hugging Face Trending chart and being dubbed by many as the fastest-growing model release in the platform's history.

Confirmed

  • 要点 According to the Hugging Face page and multiple sources, the Kimi K3 model entry was published by Unsloth, specifically identified as unsloth/Kimi-K3.
  • 要点 Page information indicates that Kimi K3 is an Image-Text-to-Text model, with related file formats including transformers and safetensors.
  • 要点 The model card also provides links to Chat, Homepage, Tech Blog, and Full Report.

Why it matters

  • 要点 The launch of Kimi K3 demonstrates extremely high community enthusiasm. Its "Image-Text-to-Text" model format also hints at Moonshot's latest exploration in multimodal capabilities, marking another significant milestone in the company's foundational model research and development.