Kimi K3 Specifications: 2.8T Params, 1M Context & KDA

Kimi K3’s spec sheet reads like a generational leap — because it is one. Here’s every major specification of Moonshot AI’s July 2026 flagship, explained in plain language. For the bigger picture, see our Kimi K3 complete guide.

Full specification table

Specification	Kimi K3
Total parameters	2.8 trillion
Architecture	KDA + Attention Residuals + Stable LatentMoE
Experts	896 routed + 2 shared, 16 routed active per token
Active parameters	~104B per token
Context window	1,048,576 tokens (1M)
Vision / video input	Native
Long-context decode speed	Up to 6.3× faster than Kimi K2
Scaling efficiency	~2.5× better than Kimi K2
Weights license	Custom Kimi K3 License — released July 27, 2026
API model name	kimi-k3

What the key specs actually mean

2.8 trillion parameters (MoE)

K3 is a mixture-of-experts (MoE) model. It doesn’t activate all 2.8T parameters for every token — instead, a router sends each token to just 16 of its 896 specialized experts (plus 2 always-on shared experts), for ~104B active parameters per token. The result: the knowledge capacity of a massive model, with the inference cost of a much smaller one. This is why K3 can compete with GPT-5.6 Sol on benchmarks while costing about half as much per task.

1M-token context window

1,048,576 tokens is roughly 750,000+ words — entire codebases, full books, or long video transcripts in a single prompt. Context size alone isn’t new, but K3 pairs it with speed (below), which is what makes 1M tokens practical rather than a marketing number.

Kimi Delta Attention (KDA)

KDA is K3’s new attention mechanism, and it’s the reason long contexts are fast. Traditional attention gets slower quadratically as context grows. KDA changes how the model attends across long sequences, delivering up to 6.3× faster decoding at long contexts compared to K2 — so a 500K-token prompt doesn’t mean waiting minutes per response.

Attention Residuals

The second architectural breakthrough. Attention Residuals improve how information flows through the model’s layers, which Moonshot credits for K3’s ~2.5× better scaling efficiency — meaning K3 extracts far more capability per unit of training compute than K2 did.

Stable LatentMoE

The routing system that keeps 896 experts training and serving stably. More experts means more specialization (better quality), but historically harder training stability; Stable LatentMoE is Moonshot’s solution, allowing the jump from K2’s 384 experts to 896.

Native vision and video input

Unlike K2, K3 accepts images and video natively — no separate adapter model. This shows up in K3’s visual-agent benchmark results, like its #1 BrowseComp score (91.2).

How K3’s specs compare to K2

Spec	Kimi K2	Kimi K3
Total parameters	1.0T	2.8T
Experts	384	896 (16 active)
Context window	256K	1M
Vision / video	—	Native
Frontend Code Arena	#18 (K2.6)	#1

We break down which model wins for which use case in Kimi K3 vs Kimi K2: Full Comparison.

The two K3 variants: k3 vs k3-256k

Kimi Code memberships offer K3 under two model IDs (the pay-as-you-go API uses a single ID, kimi-k3):

Model ID	Context	Quota cost	Notes
k3	Up to 1M on Pro+ (legacy Allegretto+); 256K on Plus	~2× baseline	Full multimodal — text, image, video
k3-256k	256K	Baseline (~half of k3)	Same results within 256K; no video input

Both are available on Plus and above (legacy Moderato and above) under the plans Moonshot relaunched on September 18, 2026, and both support low / high / max thinking effort levels. For everyday Q&A, code completion, and single-file edits, k3-256k is the cost-efficient default; reach for the 1M k3 when you need whole-codebase context or video understanding. Tier details are in our Kimi Code K3 setup guide; note that the kimi-for-coding ID in the same catalog now serves K2.8 Preview, not K3.

Can you self-host these specs?

Honestly — not easily. 2.8 trillion parameters, even in MoE form with 16 active experts, requires serious multi-GPU infrastructure. The weights were released on July 27, 2026 on Hugging Face (moonshotai/Kimi-K3) under the custom Kimi K3 License, and at ~1.56TB in native MXFP4 they are realistic for well-resourced labs and companies, not hobbyists (open weights guide). Unsloth GGUFs (~466GB and up) and disk-streaming experiments exist, but they run at roughly 0.1 tokens per second or slower (8GB RAM experiment). Everyone else should use the Kimi app, Kimi Work desktop app, or the API (docs at platform.kimi.ai, base URL https://api.moonshot.ai/v1).

FAQ

How many parameters does Kimi K3 have? 2.8 trillion total parameters in a mixture-of-experts design, with 16 of 896 routed experts (plus 2 shared experts) active per token — roughly 104B active parameters per token.

What is Kimi Delta Attention (KDA)? K3’s new attention mechanism — a linear-attention variant enabling up to 6.3× faster long-context decoding than K2, with a prefill-cache implementation contributed to vLLM.

What is Kimi K3’s context window? 1,048,576 tokens (1M) — entire codebases, books, or long video transcripts in one prompt.

Can I self-host Kimi K3? Yes, but not cheaply. The weights were released July 27, 2026 on Hugging Face (moonshotai/Kimi-K3) under the custom Kimi K3 License, in native MXFP4 (~1.56TB), and Moonshot recommends supernode deployments with 64+ accelerators. Most users should use the app or API.

The takeaway

Kimi K3’s specifications aren’t just bigger numbers — the KDA + Attention Residuals architecture makes its headline features (1M context, 896 experts) usable in the real world. At launch it ranked #4 of 189 models on Artificial Analysis’s previous index version; on the current Intelligence Index (v4.3.2) it scores 44, #3 of 115 large open-weight models. Want to see what those specs deliver in practice? Start with the complete Kimi K3 guide. The Kimi app’s free tier is fine for a first look, but confirmed K3 access (and the 1M context) requires a paid plan.