Kimi K3

Moonshot AI’s open-weight, natively multimodal agentic model, released 2026-07-28 (weights on Hugging Face, repo MoonshotAI/Kimi-K3). Successor to Kimi K2.6 (~1.1T). This page was a T4 placeholder built on one Medium essay until the release landed; it is now sourced from the model’s own repository and technical report.

Architecture

• 2.8T total parameters, 104B activated per token — sparse MoE, 16 of 896 experts live.
• 93 layers: 69 using Kimi Delta Attention, 24 Gated MLA, 1 dense. 96 attention heads, hidden dimension 7,168, 160K vocabulary.
• 1,048,576-token context (1M).
• MoonViT-V2 vision encoder (401M) — multimodal is native, not bolted on.
• Native MXFP4 weights with MXFP8 activations, applied as quantization-aware training rather than post-hoc. Relevant to quantization: the model ships quantized by design, so the usual precision-versus-quality trade is made upstream by the lab.
• Thinking is always on, with reasoning effort low / high / max and preserved thinking history across turns. (Compare sakana-fugu, whose dial starts at high, and claude-opus-5‘s low→max.)

Benchmarks (vendor-run, 2026-07-28)

All Kimi K3 numbers at effort max, temperature 1.0. Selected rows against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2:

The methodology notes are the interesting part. This is a vendor benchmark table that documents its own confounds, which is rare enough in this corpus to record:

• Harnesses differ per model. K3 runs in Kimi Code; rivals run in Claude Code, Codex or Terminus, with “best score across harnesses” taken for others. On DeepSWE the harness alone moves K3 from 67.5 (Kimi Code) to 67.3 (mini-SWE-agent) — small here, but disclosed.
• SWE-Marathon used a modified branch: an H20-calibrated fork of the official tasks from before the v1.1 release, with Docker images, performance gates and reference oracles recalibrated, validators unchanged.
• They caveat their own win: “Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance.”
• BrowseComp needed context management: 91.2 with compaction triggered at 300K, 90.4 using the full 1M window without it — so the headline 1M context is not free to use.

Still first-party and still selective, but the disclosure standard sits above Anthropic’s ratio-only Opus 5 claims. Treat it as a serious self-report and wait for a neutral referee — artificial-analysis or arc-prize — before ranking it.

The license is the strategic news

Weights and code ship under a bespoke Kimi K3 License, not the Modified MIT of K2.6. It reads MIT-shaped — use, copy, modify, distribute, sublicense, sell, fine-tune, deploy — with one carve-out: if you run a Model-as-a-Service business and your revenue exceeds US$20M over any consecutive 12 months, you must sign a separate agreement with Moonshot before commercial use. Hosting for end-user products or relaying to models hosted by others is explicitly excluded from the definition.

That is a revenue-gated commercial-hosting clause, structurally the same move as Meta’s non-OSI llama license: free for essentially everyone except the clouds and API resellers who would monetize serving it. The open-weight-models licensing axis now has three shapes — true Apache-2.0/MIT (qwen, glm-52, inkling), Meta’s user-threshold license, and Moonshot’s revenue-threshold-on-serving license.

Availability

Weights on Hugging Face; API at platform.kimi.ai with OpenAI- and Anthropic-compatible endpoints (the compatibility-as-strategy pattern in synthesis); deployment recipes for vLLM, SGLang and TokenSpeed/Lightseek. No pricing in the repository.

Moonshot’s own agent product doesn’t run it yet. kimi-claw — the managed OpenClaw deployment Moonshot launched a day later — ships K2.6 Thinking, not K3. No reason is given, and the obvious guesses (a 24/7 agent has a different cost profile; productizing a day-old 2.8T model takes longer than announcing it) aren’t sourced. Recorded because the gap between a lab’s flagship and the model its own products run is a fact the rate card never shows.

What the T4 placeholder got right and wrong

Worth keeping as a calibration record for fable5-gpt56-kimi-k3-creators, the Medium essay this page was built on. Right: 2.8T parameters, open weights, the 2026-07-27 date (repo live on the 28th). Not established: “#1 for editorial writing” appears nowhere in the release, and the claim that K3 edged Fable 5 on coding is half true at best — K3 leads on SWE-Marathon (42.0 vs 35.0) and trails on DeepSWE (67.5 vs 70.0). A weak source got the checkable specs right and the rankings wrong, which is roughly what you would expect and worth remembering the next time a T4 anchor is the only thing available.

Why it matters

An open-weight model trading wins with the closed frontier, released with its weights, is the sharpest data point yet for the “does the frontier premium survive?” question in synthesis — harder evidence than glm-52 or agents-a1 because the comparison set is the current flagships, not last generation’s. The counterweight is the license: the frontier is closing from below, but the terms are getting less open as it does.

Related

open-weight-models · claude-fable-5 · claude-opus-5 · glm-52 · llama · quantization · llm-benchmarks · kimi-claw · fable5-gpt56-kimi-k3-creators · synthesis