Kimi——From a 200K Long-Context Tool to a 2.8T Open-Source Frontier, and K3's Architectural Leap

🌏 繁體中文版

In 2023, while the world chased OpenAI, a Chinese startup called Moonshot AI (月之暗面) made a different bet — instead of rushing a "Chinese ChatGPT," it wagered on the ultra-long context window. In July 2026 that path reached its extreme: Kimi K3 — the world's first open-source 3T-class model, 2.8T parameters, 104B active, 1M context, scoring 60 on the Artificial Analysis Intelligence Index, tied with GLM-5.3 for open-source #1. This post traces Kimi's full evolution from long-context tool to open-source frontier, and K3's architectural leap.

For how to read the benchmark numbers cited here, see the AI Model Evaluation Sources Guide. This is the ninth family deep-dive in the AI Model Landscape Overview series.

Family Evolution Timeline

Three years, eight generations. The first act was "long-context differentiation"; the second is scale and architecture advancing on two tracks at once — K3 pushed total params to 2.8T while lifting efficiency 2.5x with a new architecture.

Two Product Lines: Open Weights and API

To read Kimi's 2026 moves, split it into two lines:

Open-source line (HuggingFace MoonshotAI): K2.5 under Modified MIT, K3 under the Kimi K3 License (custom, see License Trap below). Weights fully open for download, fine-tune, self-host. This line owns ecosystem and research influence — K3 is the first open 3T-class model ever, meaningful for opening frontier scale to the community.

Commercial line (platform.kimi.ai API): K3 API priced at $3.00 / $15.00 per 1M tokens (cache-hit input $0.30), backed by the Mooncake disaggregated inference architecture with >90% cache hit on coding workloads. This line owns revenue.

One important date: K2.5 and the moonshot-v1 series sunset entirely on 2026-08-31, and new registrations can no longer select them. Kimi's product重心 has fully shifted to K3.

Architecture: Why 2.8T Achieves 2.5x Efficiency

Kimi Delta Attention (KDA)

K3's core architectural innovation. KDA provides efficient long-sequence mixing with periodically interleaved Gated MLA layers to preserve global interaction. It solves the fundamental problem of "how information flows effectively across 1M context" — traditional attention dilutes information at that length; KDA keeps the model effective on both sequence length and model depth.

Attention Residuals (AttnRes)

Lets each layer selectively attend to representations of all preceding layers, not just the previous one. This improves information flow across model depth — a key reason K3 stays stable at 93 layers.

Stable LatentMoE

K3 expands the routed expert space to 896, activating 16 + 2 shared per token. With normalization, SiTU-GLU, and Quantile Balancing, it stabilizes optimization at extreme sparsity (~56:1). Versus K2's 384 experts / 8 active, K3 doubles both expert count and active ratio.

MoonViT-V2: Vision Encoder Trained From Scratch

K3's vision encoder MoonViT-V2 (401M params) is trained entirely from scratch with next-token prediction, no SigLIP initialization. Moonshot found that pretrained vision encoders (e.g., SigLIP) attached to an LLM show unstable gradients and frequent spikes; MoonViT-V2 stays stable throughout and matches the baseline on vision eval — proving large multimodal models "don't need contrastive pretraining for initialization."

Kimi K3: How to Choose

The K3 generation offers open weights and API; K2.5 is retiring:

Pricing and specs from the Kimi K3 official blog and Kimi API model list.

License Trap: Kimi K3 License Is Not MIT

K2.5 shipped under Modified MIT (relatively permissive), but K3 switched to the custom Kimi K3 License — the same signal as Qwen3.8-Max's custom terms: flagship open weights, but the license is no longer the most permissive tier. Specific restrictions on commercial deployment and redistribution need line-by-line confirmation.

The but: "open 3T-class" is real, but "freely commercial" depends on license details. If your deployment needs license certainty, K3's Kimi K3 License is less clean than GLM's MIT or most of Qwen's Apache 2.0. For the most permissive license at large scale, GLM-5.3 (MIT) remains the safer pick.

Performance Position

Three honest buts: K3 ties GLM-5.3 at 60 on the Intelligence Index (open-source #1), but overall still trails Claude Fable 5 and GPT-5.6 Sol; the gap is clear on HLE-level difficulty; even quantized, 2.8T weights need a multi-node datacenter — individual developers can't actually run it without the API.

Sub-lines and Ecosystem: A Table of Kimi's Model Range

Beyond the general line, Moonshot runs several sub-lines:

Two trends:

Capability consolidates into the mainline. Same script as Qwen, DeepSeek, GLM — Moonshot folds specialist lines back into K: K3 is natively multimodal, natively coding, no separate sub-line needed.

Open license stratifies. K2.5 was Modified MIT; K3 switched to the custom Kimi K3 License. Like Qwen and Llama 4, "open-source" is becoming different things at different model tiers — small models permissive, flagship conditional.

Position Against Competitors

Place K3 in the August 2026 open-source landscape:

• vs GLM-5.3 (744B MIT): both tied at 60 on the Intelligence Index (open-source #1). K3 is bigger (2.8T vs 744B), but GLM's MIT license is cleaner; Kimi's edge is 1M context and architectural efficiency
• vs Qwen3.8-Max (2.4T): Qwen spans a wider size spectrum (0.8B–2.4T), but K3 is the only open 3T-class option; both flagships use custom licenses
• vs DeepSeek V4: DeepSeek's price ($0.28/$0.42) is an order of magnitude below K3's $3/$15, and MIT-licensed; K3's edge is scale and long context
• vs Claude / GPT frontier: K3 still trails Fable 5 and GPT-5.6 Sol overall, but open deployment and 1M context are what closed models can't offer

What This Means for Agent Developers

• Need open 3T-class model → Kimi K3: world's first, 2.8T params + 1M context, currently the only option
• Complex coding agent → K3: DeepSWE 67.5, ProgramBench near closed, with Kimi Code CLI
• Ultra-long docs / knowledge work → K3's 1M context, fits research reports, long-doc analysis, multi-step reasoning
• Need cheapest open API → DeepSeek V4 Flash ($0.28/$0.42) is an order of magnitude cheaper than K3 ($3/$15)
• Need license certainty → GLM-5.3 (MIT) is cleaner than K3's Kimi K3 License
• Need strongest coding → Claude Opus 5 still leads SWE-bench

Overall

Kimi's story is "an overlooked direction grew an open-source frontier." From the 200K long-context chatbot of 2023 to K3 of July 2026 — 2.8T params, 104B active, 1M context, the world's first open 3T-class model. In three years it went from "Chinese long-context tool" to "open-source scale benchmark."

K3's architectural upgrades (KDA + AttnRes + Stable LatentMoE) delivered a 2.5x scaling-efficiency gain. At 60 on the Intelligence Index, tied with GLM-5.3 for open-source #1, it reaches frontier level on coding, knowledge work, and reasoning.

What's worth remembering is the license: on the "open-source" spectrum, K3 sits in the "downloadable but conditional" middle — the Kimi K3 License is less clean than GLM's MIT or most of Qwen's Apache 2.0. For developers needing 3T-class scale and able to accept a custom license, K3 is currently the only open option; for those needing license certainty, GLM-5.3 remains the safer pick.

References

• Kimi K3: Open Frontier Intelligence — Official Blog
• Kimi K3 — arXiv:2607.24653
• MoonshotAI/Kimi-K3 — Hugging Face
• Kimi API Model List
• Kimi K1.5 Technical Report
• Kimi K2 Technical Report
• Kimi K3 — kimi.ai/ai-models
• AI Model Evaluation Sources Guide — this site
• AI Model Landscape Overview — this site