clem 🤗

Jared Palmer

The New Stack

<p>Moonshot AI released the full weights of Kimi K3 on July 27, 2026, making it the largest open-weight model ever released. At 2.8 trillion total parameters, it’s a significant milestone — though the gap between “free to download” and “actually usable” is worth understanding before getting too excited.</p>
<h2>What Kimi K3 is</h2>
<p>Kimi K3 is a Mixture-of-Experts model that activates only 104 billion of its 2.8 trillion parameters per token. It supports a 1 million token context window, native multimodal input (text, images, and video), and an OpenAI/Anthropic-compatible API. The architecture introduces a new attention mechanism called Kimi Delta Attention, which Moonshot claims makes long-context inference up to six times cheaper than previous approaches.</p>
<p>The weights ship under a Modified MIT license, permissive enough for commercial use. Training data and training code aren’t included, so this is open-weight rather than fully open-source — but for most teams, the license is what actually matters.</p>
<h2>The benchmarks</h2>
<p>On the Artificial Analysis Intelligence Index, Kimi K3 scores 57, behind Claude Fable 5 (60) and GPT-5.6 Sol (59), but ahead of Claude Opus 4.8 (56). On the Frontend Code Arena it ranks first overall with an Elo of 1,679. Cognition reports it’s the first open-source model they’ve tested that passes their FrontierCode 1.1 benchmark, and it’s now available in Devin Desktop and CLI.</p>
<p>Moonshot says it roughly matches Claude Fable 5 on coding tasks at about a third of the cost per token — though it runs about four times slower.</p>
<h2>The infrastructure problem</h2>
<p>The full weights come in at roughly 1.4 terabytes using MXFP4 quantization. Self-hosting requires around 64 NVIDIA H100 or B200 GPUs across eight servers. For most teams, that’s not a realistic option. Self-hosting only makes economic sense at high inference volume where per-token savings outweigh the infrastructure cost. Everyone else will use an inference provider.</p>
<p>Moonshot is also candid about some limitations: the model uses always-on reasoning and can be over-proactive on ambiguous prompts.</p>
<h2>The geopolitical context</h2>
<p>The release isn’t without controversy. Anthropic and U.S. officials have alleged that Moonshot distilled outputs from American frontier models during training. Moonshot denies this. Enterprise buyers will need to weigh that alongside the benchmark numbers and infrastructure costs — compliance and regulatory risk are real considerations depending on the use case.</p>
<h2>Why this matters</h2>
<p>This release lands the same week as DeepSeek V4’s stable release on July 24, making late July 2026 an unusually dense period for open-weight releases. It also coincides with an open-weight AI policy letter signed by Nvidia, Meta, Microsoft, and others — the timing probably isn’t accidental.</p>
<p>A model scoring within three points of Claude Fable 5 on a major intelligence index is now free to download. That doesn’t collapse the closed-model business overnight — the infrastructure barrier is real. But it does make it harder to argue that frontier-level AI capability requires a frontier-level budget.</p>


Collections

Moonshot AI has released the full weights of Kimi K3, a 2.8 trillion parameter mixture-of-experts model — the largest open-weight model by parameter count to date. Released under a Modified MIT license permitting commercial use, the weights total ~1.4TB in MXFP4 quantization. Kimi K3 scores 57 on the Artificial Analysis Intelligence Index (behind Claude Fable 5 at 60 and GPT-5.6 Sol at 59) and ranks first on the Frontend Code Arena with 1,679 Elo. Moonshot also introduced Kimi Delta Attention, claiming up to 6x cheaper long-context inference. The model is already integrated into Devin Desktop and CLI. While self-hosting requires significant multi-GPU infrastructure, the release narrows the gap between open-weight and closed frontier models, following DeepSeek V4's stable release just days earlier.

Moonshot AI releases Kimi K3 weights: a 2.8T open-weight model that tops the Frontend Code Arena

<p>Amazing 😍 ✌️ ❤️</p>
<p>Well done China!</p>


<p>“The infrastructure barrier is real.”</p>
