Moonshot releases Kimi K3 open weights

• Kimi K3 is now public as open weights: _akhaliq's launch post points to the Hugging Face repo, and icreatelife framed it as an open-weight model near frontier performance.
• The headline spec is huge but sparse: the linked repo lists 2.8T total parameters, 104B active parameters, 896 experts, 16 selected experts per token, and a 1,048,576-token context window.
• Creator workflows showed up immediately: _akhaliq's Claude Code demo put K3 behind Hugging Face's Claude Code picker, while Everlier's Cursor test sent it at a staging-API pentest prompt.
• Local ownership still means serious hardware: LLMJunky's VRAM joke got corrected to "<3tb" in LLMJunky's later reply, and another LLMJunky reply called it ultra-wealthy or enterprise territory.
• Moonshot is treating openness as distribution: aakashgupta claimed daily revenue grew at least 6x after K3 launched and ARR reached $300M in June.

Moonshot's Hugging Face model card buries a real workflow gotcha: K3 returns reasoning_content, and multi-turn apps are expected to pass that reasoning history back as state. The Hugging Face Claude Code guide shows why the first demos appeared so fast: Claude Code can be pointed at HF's router with environment variables or the hf-claude picker. Baseten's launch writeup says its K3 replica uses eight GB300 GPUs, while Modal claimed 460 tokens per second on release day.

Moonshot's model card calls Kimi K3 an open-weight, native multimodal agentic model with 2.8T total parameters, 104B active parameters, 93 layers, and a 1,048,576-token context window.

The useful spec sheet is sparse enough to scan:

• Architecture: Mixture-of-Experts.
• Attention: 69 KDA layers plus 24 Gated MLA layers.
• Experts: 896 total, 16 selected per token.
• Vision encoder: MoonViT-V2, 401M parameters.
• Quantization: MXFP4 weights and MXFP8 activations.
• Vocabulary: 160K.
• Context length: 1,048,576 tokens.

One wrinkle for video people: Moonshot's feature copy says K3 understands text, images, and video within the same model, while the summary table's modality row says Text, Image.

Moonshot's official table does not show a clean sweep. It shows a model trading blows with hosted frontier systems across coding, agentic, and vision tasks.

K3 leads the listed comparison set on several agentic or creator-adjacent rows:

• BrowseComp: 91.2, ahead of Claude Fable 5 at 88.0 and GPT-5.6 Sol at 90.4.
• MCPMark-Verified: 94.5, ahead of Claude Fable 5 at 87.4 and GPT-5.6 Sol at 92.9.
• AutomationBench: 30.8, ahead of Claude Fable 5 at 29.1 and GPT-5.6 Sol at 29.7.
• SpreadsheetBench 2: 34.8, essentially tied with Claude Fable 5 at 34.7 and ahead of GPT-5.6 Sol at 32.4.
• OmniDocBench: 91.1, ahead of Claude Fable 5 at 89.8 and GPT-5.6 Sol at 85.8.

The same table has K3 behind Claude Fable 5 on DeepSWE, FrontierSWE, OfficeQA Pro, and Legal Research Bench. Frontier open weights with visible tradeoffs is the interesting part.

K3 is not just a chat-completions swap if an app wants multi-turn tool use. Moonshot's model card says K3 always has thinking enabled, returns reasoning_content, and supports reasoning_effort values of low, high, and max, with max as the default.

For multi-turn conversations and tool calls, the same card says the complete assistant message must be passed back into messages as-is, including reasoning_content and tool_calls, not only content.

That is the migration gotcha for vibe-coded agents: the reasoning trace becomes part of the model's working state.

The first creator workflow was a terminal picker. _akhaliq's screenshot showed moonshotai/Kimi-K3 highlighted inside Hugging Face's Claude Code model selector, with 129 models available in the picker.

Hugging Face's Claude Code integration guide describes two setup paths:

• hf-claude, an HF CLI extension with an interactive model and provider picker.
• Manual environment variables, including ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY, default Opus/Sonnet/Haiku model slots, and CLAUDE_CODE_SUBAGENT_MODEL.
• Provider suffixes and routing policies, including :cheapest, :fastest, and :preferred.

In Cursor, Everlier used K3 for a timeboxed staging-API pentest prompt that asked for valid findings only when there was an end-to-end exploit path toward unauthorized data access.

The 47 TB line was the meme, not the deployment spec. LLMJunky's correction put the range at 1.3 to 2.3 TB, another correction shortened it to "< 3tb," and a later reply said it was already four bit.

The hardware question underneath the jokes was real. LLMJunky noticed tp8 and bf16 in an AMD/SGLang context and asked how K3 was working on 8× MI350X, roughly 2.3 TB of VRAM.

Baseten's day-zero API writeup gives one production answer: native MXFP4 weights with MXFP8 activations, eight NVIDIA GB300 GPUs per replica, and KV-aware routing to reduce repeated prefill.

Hosted K3 access appeared alongside the weights. Everlier immediately wanted other frontier models running around 90 tokens per second, which is the right bar for whether an open-weight monster feels interactive.

Modal's release-day post claimed K3 at 460 tokens per second, up to 1.5 million TPM per GPU, token-based pricing through its Shared API, and dedicated capacity through Auto Endpoint.

Baseten said K3 was live on its Model APIs with vision input and the full 1M-token context window, after early work with Moonshot, Inferact, RadixArk, vLLM, and SGLang.

The weights are open, but the license is not a plain Apache-style release. Unite.AI's license breakdown says the Kimi K3 License allows use, copying, modification, distribution, sublicensing, sale, running, deployment, and fine-tuning for most users.

The same breakdown says a separate Moonshot contract is required when a Model-as-a-Service business provides third-party inference or fine-tuning and exceeds $20 million in revenue over 12 months.

It also says products with more than 100 million monthly active users or more than $20 million in monthly revenue must prominently display "Kimi K3" in the interface.

aakashgupta argued that the open release works as distribution: every self-hosted deployment becomes a trial, and the paid API stays OpenAI SDK compatible enough to migrate with one config change.

A CNBC TV18 report based on Bloomberg also said Moonshot's daily sales grew at least sixfold after K3's release and that the company was seeking a funding round at a $50B valuation.

The deployment economics landed in the community as a different kind of number. kaigani wrote that you could "essentially build Kimi K3 for a full-time developer's annual salary," while LLMJunky joked the ask was "$450,000 in compute."

Hacker News had the predictable open-model pileup, with the main Kimi-K3 thread drawing more than 500 comments around hardware, hosting, licensing, and whether a 3T-class open-weight model changes the floor for everyone else.

The creator-content loop started immediately too. gokayfem posted a fully AI-generated 10-minute course on Kimi K3, attached to a thread that also showed a Flux 3 prompt for "Daily life of the Aztecs."

K3's model card explicitly names interactive visualizations, widgets, dashboards, motion design, and video editing as target knowledge-work outputs. That made the course demo feel less like launch-day fan art and more like the first marketing asset made in the same culture the model is trying to serve.