Kimi K3.1 Pro (2.8T MoE) Architecture Parameters, Benchmarks & VRAM
Moonshot October 2026 flagship 2.78T MoE (93 layers, 896 experts) with 1M input & 1M output context.
🧬 Kimi K3.1 Pro (2.8T MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (moonshotai/Kimi-K3.1-Pro) config.json & Model Card
1. Backbone & Dimensions
kimi_k31- Total / Active Params
- 2.78T / 104B
- Hidden Size (d_model)
- 7,168
- FFN Intermediate Size
- 33,792
- MoE Expert Dim
- 3,072
- MoE Experts
- 896 Routed · Top-50 + 2 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=0.00001
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 93 Layers
- Hidden Layer Layout
- 93-Layer Sparse MoE (896 Experts, Top-50 Active)
- Attention Heads (Q / KV)
- 96 Q / 96 KV
- Head Dimension
- 75
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 1,048,576 tokens
- RoPE Configuration
- RoPE (原生 1,048,576 Context)
- Vocab Size
- 163,840 (Untied)
📋 Kimi K3.1 Pro (2.8T MoE) Official API & System Card Specifications
Verified against official Moonshot Kimi developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- kimi-k3.1-pro
- Release Date
- 2026-10-23
- Weight License
- Open-Weight (API Spec)
- Native Reasoning
- ✅ Supported
- Function Calling & JSON Schema
- ✅ 原生支持
- Batch API 异步批处理
- ✅ Supported (50% off)
2. Context, Output & Modalities
Limits- Max Context Window
- 1,024,000 tokens (1024K)
- Max Output Tokens
- 1,049,000 tokens
- Input Modalities
- Text + Image + Video
- Output Modalities
- Text
- Direct China Access
- ✅ Direct CN + RMB
- Deployment Mode
- Open-Weight + Cloud API
3. Official Pricing & Cache
CNY / 1M- Input / 1M Tokens
- ¥ 13
- Cached Input / 1M
- ¥ 1.3 -90%
- Output / 1M Tokens
- ¥ 68
- Tiered Pricing Notes
- Moonshot AI: Moonshot AI official CN platform with auto prefix cache
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Kimi K3.1 Pro (2.8T MoE) Full Benchmark Scorecard
30 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 10 items
BenchCAD variant evaluated with access to a Python tool for programmatic CAD reasoning.
Kimi Code Bench v2 is Moonshot AI's in-house benchmark for evaluating coding agents on realistic software engineering tasks across 10+ mainstream programming languages and a production tech stack spanning backend services, infrastructure, performance engineering, systems…
DeepSWE is a software engineering agent benchmark evaluated with the mini-swe-agent harness, where each task is solved in an isolated container with no internet access. It measures an agent's ability to autonomously resolve real-world coding issues end to end.
ExploitBench is a cybersecurity benchmark that evaluates a model's ability to discover and exploit software vulnerabilities, reported as the fraction of challenges where the model captures the target (Cap%).
BenchCAD is a benchmark for programmatic CAD reasoning built from 17,900 execution-verified CadQuery programs spanning 106 industrial part families, roughly half anchored to real ISO, DIN, EN, ASME, and IEC specification tables.
Terminal-Bench 4.0 evaluates AI agents on real-world work in terminal and command-line containerized environments across 66 tasks, with emphasis on science-adjacent and frontier engineering problems.
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
Terminal-Bench-Science 0.1 is a Stanford-led community benchmark of 70 tasks drawn from scientific research workflows across the life, physical, earth, mathematical, and engineering sciences.
PostTrainBench evaluates a model's ability to autonomously post-train base models. Given pretrain-only base models, the agent must complete the full pipeline of data synthesis, training, evaluation, and iteration within a time budget, scored across downstream benchmarks such as…
ExploitGym is a large-scale, realistic benchmark built from real-world vulnerabilities across userspace programs, Google's V8 engine, and the Linux kernel.
🤖 Agents & Tools 8 items
BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled information.
AutomationBench-AA is Artificial Analysis's independently run version of AutomationBench, covering 657 real-world SaaS workflow tasks across 40 simulated applications (e.g. Gmail, Slack, Salesforce, HubSpot).
OfficeQA Pro evaluates AI models on professional knowledge-work questions and tasks drawn from real office workflows, including document analysis, spreadsheet reasoning, and information synthesis across business domains.
APEX-Agents is a benchmark evaluating AI agents on long horizon professional tasks that require sustained reasoning, planning, and execution across complex multi-step workflows.
Version 2.1 of the Artificial Analysis GDPval evaluation of professional knowledge work, reported as an Elo rating.
AA-Briefcase v1.1 evaluates multi-hour professional office work and reports Elo scores. Kept separate from unversioned AA-Briefcase results.
Agents' Last Exam is a challenging benchmark for AI agents on hard, long-horizon tasks that test sustained reasoning, planning, and tool use, reported with and without tool access.
Harvey's Legal Agent Benchmark evaluates AI agents on complex legal workflows and reports the all-pass rate.
📐 Math 2 items
Full multimodal Humanity's Last Exam evaluated with tool use enabled, distinct from its text-only subset.
Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions.
👁️ Multimodal 8 items
A benchmark for early-stage visual reasoning and perception on child-like vision tasks.
Chartography evaluates professional chart understanding with 100 expert-written tasks drawn from real charts across science, engineering, finance, manufacturing, supply chain, and healthcare.
MMMU-Pro variant evaluated with tool access enabled.
CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.
A more robust multi-discipline multimodal understanding benchmark that enhances MMMU through a three-step process: filtering text-only answerable questions, augmenting candidate options, and introducing vision-only input settings.
PerceptionBench is Moonshot AI's internal benchmark for evaluating atomic visual perception capabilities.
GDP.pdf is a knowledge-work vision benchmark that evaluates models on economically valuable professional tasks presented as visual documents (PDFs), testing document-based reasoning, chart and table interpretation, and problem solving without tools.
ZEROBench is a challenging vision benchmark designed to test models on zero-shot visual understanding tasks.
🧠 Reasoning & Science 2 items
A challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. Questions are Google-proof and extremely difficult, with PhD experts reaching 65% accuracy.
BioMysteryBench evaluates a model's ability to reason through challenging molecular biology problems, reporting performance on a hard subset and on the subset of problems solved by human experts.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ Kimi K3.1 Pro (2.8T MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 1830.5%
Exceeds usable VRAM by 1628.06 GB
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run kimi-k3-1-pro-2800bInsufficient Local VRAM? Ways to Run Kimi K3.1 Pro (2.8T MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 Kimi K3.1 Pro (2.8T MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 5560 GB | 1.76 GB | 5674.26 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 2953.75 GB | 1.76 GB | 3015.88 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 2279.6 GB | 1.76 GB | 2328.25 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 1977.28 GB | 1.76 GB | 2019.88 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 1685.37 GB | 1.76 GB | 1722.14 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 1476.88 GB | 1.76 GB | 1509.47 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 1358.73 GB | 1.76 GB | 1388.95 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 1094.63 GB | 1.76 GB | 1119.57 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for Kimi K3.1 Pro (2.8T MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of Kimi K3.1 Pro (2.8T MoE)?
Kimi K3.1 Pro (2.8T MoE) uses KimiK31ForConditionalGeneration with 93 layers (93-Layer Sparse MoE (896 Experts, Top-50 Active)), hidden_size=7168, 96 Q heads / 96 KV heads (head_dim=75), and vocab_size=163,840.
How much VRAM does Kimi K3.1 Pro (2.8T MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), Kimi K3.1 Pro (2.8T MoE) requires 1722.14 GB VRAM. Full FP16 requires 5674.26 GB.