Kimi K2.5 (1T MoE) Architecture Parameters, Benchmarks & VRAM
1T total / 32B active multimodal MoE supporting Agent Swarm orchestration.
🧬 Kimi K2.5 (1T MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (moonshotai/Kimi-K2.5) config.json & Model Card
1. Backbone & Dimensions
kimi_k25- Total / Active Params
- 1.026T / 32B
- Hidden Size (d_model)
- 7,168
- FFN Intermediate Size
- 18,432
- MoE Expert Dim
- 2,048
- MoE Experts
- 384 Routed · Top-8 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=0.00001
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 61 Layers
- Hidden Layer Layout
- 61-Layer Sparse MoE (384 Experts, Top-8 Active)
- Attention Heads (Q / KV)
- 64 Q / 64 KV
- Head Dimension
- 112
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 262,144 tokens
- RoPE Configuration
- RoPE (原生 262,144 Context)
- Vocab Size
- 163,840 (Untied)
📋 Kimi K2.5 (1T MoE) Official API & System Card Specifications
Verified against official Moonshot Kimi developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- kimi-k2.5
- Release Date
- 2026-01-27
- Weight License
- Open-Weight (API Spec)
- Native Reasoning
- ✅ Supported
- Function Calling & JSON Schema
- 部分支持
- Batch API 异步批处理
- ✅ Supported (50% off)
2. Context, Output & Modalities
Limits- Max Context Window
- 256,000 tokens (256K)
- Max Output Tokens
- 33,000 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
- ¥ 4
- Cached Input / 1M
- ¥ 0.8 -80%
- Output / 1M Tokens
- ¥ 21
- Tiered Pricing Notes
- Alibaba Bailian: Hosted on Alibaba Cloud Bailian (CN Direct)
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Kimi K2.5 (1T MoE) Full Benchmark Scorecard
39 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 8 items
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
A multilingual benchmark for issue resolving in software engineering that covers Java, TypeScript, JavaScript, Go, Rust, C, and C++.
PaperBench is a benchmark for evaluating AI agents on their ability to replicate research papers. It tests models on complex, multi-step workflows involving code implementation, experimentation, and reproducing scientific results from academic publications.
Terminal-Bench 2.0 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.
SWE-Bench Pro is an advanced version of SWE-Bench that evaluates language models on complex, real-world software engineering tasks requiring extended reasoning and multi-step problem solving.
SciCode is a research coding benchmark curated by scientists that challenges language models to code solutions for scientific problems.
CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.
FrontierSWE measures whether an agent can complete open-ended technical projects at the scale of hours to tens of hours, spanning systems optimization, large-scale code construction, and applied ML research. Performance is reported as a dominance score, where higher is better.
🤖 Agents & Tools 4 items
WideSearch is an agentic search benchmark that evaluates models' ability to perform broad, parallel search operations across multiple sources. It tests wide-coverage information retrieval and synthesis capabilities.
DeepSearchQA is a benchmark for evaluating deep search and question-answering capabilities, testing models' ability to perform multi-hop reasoning and information retrieval across complex knowledge domains.
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.
Seal-0 is a benchmark for evaluating agentic search capabilities, testing models' ability to navigate and retrieve information using tools.
📐 Math 8 items
All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.
Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds.
MathVista-Mini is a smaller version of the MathVista benchmark that evaluates mathematical reasoning in visual contexts.
A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks.
MATH-Vision is a dataset designed to measure multimodal mathematical reasoning capabilities.
IMO-AnswerBench is a benchmark for evaluating mathematical reasoning capabilities on International Mathematical Olympiad (IMO) problems, focusing on answer generation and verification.
LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses.
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 14 items
InfoVQA test set with infographic images requiring joint reasoning over document layout, textual content, graphical elements, and data visualizations with elementary reasoning and arithmetic skills.
OCRBench: Comprehensive evaluation benchmark for assessing Optical Character Recognition (OCR) capabilities in Large Multimodal Models across text recognition, scene text VQA, and document understanding tasks.
OmniDocBench 1.5 is a comprehensive benchmark for evaluating multimodal large language models on document understanding tasks, including OCR, document parsing, information extraction, and visual question answering across diverse document types.
Video-MME is the first-ever comprehensive evaluation benchmark of Multi-modal Large Language Models (MLLMs) in video analysis.
Video-MMMU evaluates Large Multimodal Models' ability to acquire knowledge from expert-level professional videos across six disciplines through three cognitive stages: perception, comprehension, and adaptation.
MMVU (Multimodal Multi-disciplinary Video Understanding) is a benchmark for evaluating multimodal models on video understanding tasks across multiple disciplines, testing comprehension and reasoning capabilities on video content.
LongVideoBench is a question-answering benchmark featuring video-language interleaved inputs up to an hour long.
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.
CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.
LVBench is an extreme long video understanding benchmark designed to evaluate multimodal models on videos up to two hours in duration. It contains 6 major categories and 21 subcategories, with videos averaging five times longer than existing datasets.
MotionBench is a benchmark for evaluating multimodal models on motion understanding in videos, testing the ability to comprehend temporal dynamics, movement patterns, and action sequences.
WorldVQA is a benchmark designed to evaluate atomic vision-centric world knowledge. It assesses models' ability to understand and reason about visual elements representing real-world knowledge.
SimpleVQA is a visual question answering benchmark focused on simple queries.
ZEROBench is a challenging vision benchmark designed to test models on zero-shot visual understanding tasks.
📜 Long Context 2 items
🧠 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.
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ Kimi K2.5 (1T MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 677%
Exceeds usable VRAM by 542.82 GB
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run kimi-k2-5-1000bInsufficient Local VRAM? Ways to Run Kimi K2.5 (1T MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 Kimi K2.5 (1T MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 2052 GB | 1.15 GB | 2095.49 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 1090.13 GB | 1.15 GB | 1114.38 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 841.32 GB | 1.15 GB | 860.6 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 729.74 GB | 1.15 GB | 746.79 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 622.01 GB | 1.15 GB | 636.9 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 545.06 GB | 1.15 GB | 558.42 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 501.46 GB | 1.15 GB | 513.94 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 403.99 GB | 1.15 GB | 414.52 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for Kimi K2.5 (1T MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of Kimi K2.5 (1T MoE)?
Kimi K2.5 (1T MoE) uses KimiK25ForConditionalGeneration with 61 layers (61-Layer Sparse MoE (384 Experts, Top-8 Active)), hidden_size=7168, 64 Q heads / 64 KV heads (head_dim=112), and vocab_size=163,840.
How much VRAM does Kimi K2.5 (1T MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), Kimi K2.5 (1T MoE) requires 636.9 GB VRAM. Full FP16 requires 2095.49 GB.