moonshotai/Kimi-K3

TL;DR: Kimi K3 is Moonshot's open-weight frontier model: a 2.8T-parameter sparse MoE (104B active) with native vision, a 1M-token context window, and always-on reasoning, released under a custom license with full weights in MXFP4. It's interesting less as "big open model" and more as an architecture statement — a mostly-linear-attention stack (Kimi Delta Attention) with sparse full-attention layers, plus a "LatentMoE" routing scheme claimed to give ~2.5× scaling efficiency over K2 — and the benchmark table puts it at or near Claude/GPT frontier level on agentic coding.

What it is

This is not a fine-tune; it's a new pretrained base with several named architectural bets. The stack is 93 layers: 69 use Kimi Delta Attention (KDA, Moonshot's gated delta-rule linear attention line, published earlier with Kimi Linear) and 24 use gated MLA (full attention with latent KV compression) — roughly a 3:1 linear-to-full ratio, which is how a 1M-token window becomes economically plausible. On top of that: "Attention Residuals" (AttnRes) and a "Stable LatentMoE" framework routing 16 of 896 experts through a 3584-dim latent bottleneck, plus 2 shared experts. Vision comes from a 401M MoonViT-V2 encoder feeding the same model — text, image, and video understanding, no separate VLM grafting. Notably, the model was quantization-aware trained from SFT onward at MXFP4 weights / MXFP8 activations, so the released 4-bit weights are the model, not a post-hoc squeeze.

The README names the components but does not explain AttnRes or Stable LatentMoE; those details live in the linked tech report. Take the "2.5× scaling efficiency over K2" claim as a pointer, not a verified number.

The evidence

The eval table is extensive and pitched directly against closed frontier models (Claude Fable 5/Opus 4.8, GPT-5.6 Sol/5.5). Highlights: GPQA Diamond 93.5, Terminal-Bench 2.1 88.3, OSWorld-Verified 84.8, and clear leads on SWE-Marathon (42.0 vs 35–40) and MCPMark-Verified (94.5 vs 87–93). BrowseComp at 91.2 uses context compaction; without it, the raw 1M window still gets 90.4, which is itself a useful data point about long-context agents.

Read the footnotes: Kimi K3 is mostly evaluated with Moonshot's own Kimi Code harness while competitors use theirs, some competitor runs hit fallbacks/refusals, and several scores are cited from third parties at different dates. The reporting is unusually transparent about this, but harness effects are real (K3 itself drops 67.5→67.3 on DeepSWE switching harnesses) — treat exact rankings as soft.

Practicalities

What's here: full MXFP4 weights (safetensors, compressed-tensors format, custom_code), deployment recipes for vLLM, SGLang, and TokenSpeed, and a hosted API (kimi-k3 on platform.kimi.ai, OpenAI/Anthropic-compatible). What's not: no base model mentioned, license is custom ("Kimi K3 License" — check terms before commercial use), and self-hosting means ~1.4TB of weights even at 4-bit, i.e., a multi-node H200/B200-class deployment. 2.7M downloads and #28 trending suggest people are doing it anyway.

Two API quirks matter: thinking is always on (tune via reasoning_effort: low/high/max), and the model was trained with preserved thinking history — you must pass the full assistant message, including reasoning_content and tool_calls, back in multi-turn conversations, or behavior degrades. Most existing agent scaffolds silently drop reasoning content, so expect to patch yours.

Try it

Cheapest path is the API; the model card's own pattern:

import openai
client = openai.OpenAI(base_url="https://platform.kimi.ai/v1", api_key="...")

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain the delta rule in linear attention."}],
    reasoning_effort="max",
    max_tokens=4096,
)
print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)
# multi-turn: append the FULL assistant message (incl. reasoning_content) back to messages

For self-hosting, start from the vLLM/SGLang recipes linked on the model card; for agentic coding, they recommend Kimi Code CLI (/model → Kimi K3). Read the full tech report before drawing conclusions about KDA/AttnRes/LatentMoE — the card only names them.