Kimi K3

What it is. Kimi K3 is Moonshot AI’s flagship open-weight model (“Open Frontier Weights”), released 2026-07-16: a 2.8T-parameter MoE with 104B activated parameters and a 1,048,576-token context window. Why it matters. It is the only model in the China AI Hub database documented as capable of both 1M-token input and 1M-token output, making it the strongest long-form generation model tracked, and it carries native visual understanding plus always-on thinking. Key characteristics. Native vision and video input, reasoning_effort low/high/max (default max), tool calling, structured output, and agent capability. What a professional should know. Temperature is fixed at 1.0 and top_p at 0.95 — you cannot tune sampling — and the license is permissive only up to a $20M revenue / 100M MAU threshold, above which Moonshot requires a separate agreement and UI attribution.

Architecture and parameters

K3 publishes the fullest architecture card in the database: 2.8T total / 104B activated (~3.7% activation), 93 layers (1 dense), 896 experts with 16 selected + 2 shared per token, 69 KDA + 24 Gated MLA attention layers, hidden dim 7168, SiTU-GLU, a 401M-parameter MoonViT-V2 vision encoder, and MXFP4 weights with MXFP8 activations. The KDA (Kimi Delta Attention) plus Gated MLA split is Moonshot’s answer to long-context economics: linear-attention layers replace part of the dense attention stack so 1M-token context stays tractable. See Mixture-of-Experts and the MoE architectures research.

What the context window actually means

K3 is the database’s unique 1M-in / 1M-out pairing. Most 1M-input models cap output at 131K–393K; K3 can generate up to 1,048,576 tokens, which reframes what it is for: long-form document synthesis, multi-file code generation, and agent runs with very long output trajectories — not just long-input analysis. The trade-off is cost: at $15/M output tokens, a single maximal response is a material expense. See the context-window research.

Pricing implications

$3.00 input / $15.00 output per 1M tokens with automatic prefix caching ($0.30 cached input). China AI Hub analysis: the 5x input-to-output ratio is the highest in the flagship tier, reflecting the premium on generation. The $15 output price, combined with a 1M output ceiling, means K3’s worst-case single-request cost is the highest in the database — budget-sensitive deployments should cap max_completion_tokens (default 131,072) rather than rely on the ceiling. API access requires a $1 minimum top-up. See the MiniMax M3 comparison.

API, coding, and agent implications

K3 targets software engineering and deep reasoning. Vendor-reported DeepSWE (67.5) and Terminal-Bench 2.1 (88.3) are strong coding/agentic signals, and agent capability is documented. The fixed sampling parameters (temperature 1.0, top_p 0.95) mean deterministic, low-temperature outputs are not achievable through the API — a real constraint for reproducible batch jobs. max_completion_tokens defaults to 131,072 and can be set to 1,048,576. The Kimi Code agent runs on K3.

Open weights and license

Weights are on Hugging Face and ModelScope under the “Kimi K3 License”: MIT-style permissive, but operators with >$20M aggregate revenue over any 12 months must sign a separate agreement, and products over 100M MAU or $20M monthly revenue must display “Kimi K3” in the UI. This is a conditional-open license, not plain MIT — commercial self-hosters above the thresholds face obligations. See licensing explained.

Benchmark interpretation

Six vendor-reported rows: GPQA Diamond 93.5, HLE-Full 43.5 (56.0 with tools), DeepSWE 67.5, Terminal-Bench 2.1 88.3, MMMU-Pro 81.6, and Video-MME 90.0. The multimodal rows (MMMU-Pro, Video-MME) reflect K3’s native vision/video; the HLE with-tools delta again shows tool access raising the ceiling. Moonshot notes some comparison scores are cited from Artificial Analysis — a third-party source — though the K3-specific rows here are vendor-reported.

What the benchmarks do not prove

These scores do not independently verify K3 against competitors, and the modality documentation is itself inconsistent: the architecture table says Text+Image while the README, launch blog and API guide also show video input. Treat the exact input-modality boundary as uncertain. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: long-form document and code generation, deep reasoning, multimodal (vision/video) analysis, and agent workloads that benefit from 1M output headroom. Less suited: high-volume low-cost chat (priced at flagship tier), latency/reproducibility-sensitive batch work (fixed sampling), and commercial self-hosting above the license thresholds.

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	High
Structured API workflows	High
Agent orchestration	High
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates that K3 occupies the top of the flagship tier on two axes — parameter scale (2.8T) and output ceiling (1M) — but pays for both in the highest output pricing and the most constrained sampling controls. The KDA/Gated-MLA architecture and 896-expert topology are the most technically transparent open release in the collection, which is an asset for researchers even if the $15 output price limits practical reach. The fixed temperature/top_p is a genuine, under-appreciated limitation for production users expecting standard sampling controls.

What is uncertain

• The input-modality boundary is inconsistent across official sources: the architecture table says Text+Image, while the README, launch blog and API guide also list video input.
• Some comparison benchmark scores are cited from Artificial Analysis (third-party); the K3-specific rows here are vendor-reported.
• Sampling is fixed (temperature 1.0, top_p 0.95) and cannot be changed through the API.

Sources

• Kimi K3 GitHub README
• Kimi API platform — model list
• Kimi API — Chat Completions spec
• Kimi K3 launch blog

Labels used above: Official fact (from primary sources), Vendor-reported claim (benchmark scores published by Moonshot), Third-party evidence (Artificial Analysis, where cited), and China AI Hub analysis (our synthesis, always introduced as such).