Moonshot AI
Kimi K3.1 HighspeedTerbuka
Latency-optimized serving tier of Kimi K3.1 — 1M context at flash-class throughput.
Inteligensi
91.5 #13
AAQI 60.0
Kecepatan output
130 t/s
ITL 7.7 ms/tok
Harga gabungan (3:1)
$3.13 /1M
Input $1.5/M · Output $8/M
Konteks
1.0M
≈ 2,500 halaman A4 · output maks 65,536
Quality Domain Radar
6-Axis Frontier Evaluation vs Category Average & Flagship
Kimi K3.1 Highspeed
Claude Mythos Preview (#1)
Average (68)
Sumbu radar diaproksimasi dari skor kategori jika pengukuran benchmark langsung tidak tersedia — lihat Metodologi.
Kelebihan & Trade-off
Di mana ia menang
- 130 tok/s — fastest 1M-context open model
- Roughly half the Pro tier’s blended price
- Same agentic tool-use lineage as K3.1 Pro
Hal yang perlu diperhatikan
- Caps reasoning effort vs. Pro — drops on the hardest tasks
- 65K output window vs. 131K on Pro
- Still needs multi-node serving to self-host
Kasus Penggunaan Killer
Real-time agent loops
High-throughput tool calling for interactive copilots.
High-volume RAG
Cheap 1M-context reads over large corpora.
Voice-adjacent LLM backends
Sub-second first token keeps spoken agents snappy.
Quickstart Developer & API
Atur
MOONSHOT_API_KEY dulu — dapatkan kunci APIcurl https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-H "content-type: application/json" \
-d '{
"model": "kimi-k3-1-highspeed",
"messages": [{"role": "user", "content": "Refactor this function for O(n) time and explain the change."}]
}'import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.moonshot.ai/v1",
api_key=os.environ["MOONSHOT_API_KEY"],
)
resp = client.chat.completions.create(
model="kimi-k3-1-highspeed",
messages=[{"role": "user", "content": "Refactor this function for O(n) time and explain the change."}],
)
print(resp.choices[0].message.content)const res = await fetch("https://api.moonshot.ai/v1/chat/completions", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.MOONSHOT_API_KEY}`,
"content-type": "application/json",
},
body: JSON.stringify({
model: "kimi-k3-1-highspeed",
messages: [{ role: "user", content: "Refactor this function for O(n) time and explain the change." }],
}),
});
console.log((await res.json()).choices[0].message.content);Sumber Daya & Tautan Resmi
Deployment API
Bandingkan host & latensi
Throughput, TTFT, dan harga cache di Groq, Cerebras, Together, DeepInfra, Fireworks, Bedrock, dan Azure.
GroqCerebrasTogether AIDeepInfraFireworks AIAWS BedrockAzure AI
Skor per sumber
Skor per kategori
Hasil benchmark
Duel Head-to-head Live
Spesifikasi Live & Arsitektur
Di balik kap
Kimi K3.1 Highspeed — 2.8T MoE, speculative decode
- Same K3.1 weights served with aggressive speculative decoding
- Sub-600ms TTFT at 130 tok/s on dedicated capacity
- 1M context retained; 65K max output
- Lower reasoning-effort ceiling than the Pro tier
Tanggal rilis
20 Okt 2026
Arsitektur
Kimi K3.1 Highspeed — 2.8T MoE, speculative decode
Parameter
2.8T MoE
Lisensi
Open Weights
Jendela konteks
1,000,000 tokens
Output maks
65,536 tokens
Modalitas
text + image
Harga cache
baca $0.15 · tulis $1.88 /1M
Penalaran
Level effort
Harga input
$1.5/M
Harga output
$8/M
Kecepatan output
130 t/s
Waktu ke token pertama
—
Data diperbarui
22 Okt 2026
Perangkat Keras & Inferensi Lokal
Data: Convly AI (CC BY 4.0) ↗GPU Min.
Multi-node cluster (e.g. 2× 8-GPU H200 141GB nodes)
Estimasi tagihan bulanan
Estimasi pengeluaran bulanan
$0.00
Menghitung penghematan vs GPT-4o…
Kutip & sematkan
Badge Markdown GitHub
Loading badge markdown...BibTeX
@misc{llmpodium2026kimik31highspeed,
title = {Benchmark Analysis and Intelligence Evaluation of Kimi K3.1 Highspeed},
author = {LLMPodium Research},
year = {2026},
url = {https://llmpodium.com/models/kimi-k3-1-highspeed}
}Sematkan Lencana Skor
Tambahkan lencana Podium Score langsung ke README GitHub atau Hugging Face Anda — menautkan ke peringkat langsung.
Pratinjau
Markdown
Loading badge snippet...HTML
Loading badge snippet...Lainnya dari Moonshot AI
Model Alternatif
Alternatif terdekat untuk Kimi K3.1 Highspeed berdasarkan kemampuan, harga, dan skor benchmark.