Kimi K2.6 (1T MoE) Architecture Parameters, Benchmarks & VRAM
1T total / 32B active open-weight multimodal MoE achieving 80.2% on SWE-Bench Verified.
🧬 Kimi K2.6 (1T MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (moonshotai/Kimi-K2.6) config.json & Model Card
1. Backbone & Dimensions
kimi_k25- Total / Active Params
- 1T / 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=1e-5
- Native Tensor Dtype
- FP8 / INT4 QAT
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 61 Layers
- Hidden Layer Layout
- 61-Layer MLA Multimodal MoE (384 Routed + 1 Shared Expert, 8 Active · 400M ViT)
- Attention Heads (Q / KV)
- 64 Q / 64 KV
- Head Dimension
- 64
- Linear / KV Compression
- MLA Low-Rank KV Compression: kv_lora_rank=512 + qk_rope_head_dim=64
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 262,144 tokens
- RoPE Configuration
- YaRN MLA RoPE (原生 262,144 Context)
- Vocab Size
- 163,840 (Untied)
- Special Token IDs
- BOS: 163584 · EOS: 163586 · PAD: 163839
📋 Kimi K2.6 (1T MoE) Official API & System Card Specifications
Verified against official Moonshot Kimi developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- kimi-k2.6
- Release Date
- 2026-04-20
- 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
- 256,000 tokens (256K)
- Max Output Tokens
- 262,000 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
- ¥ 6.5
- Cached Input / 1M
- ¥ 1.1 -83%
- Output / 1M Tokens
- ¥ 27
- Tiered Pricing Notes
- Moonshot AI: Moonshot AI official CN platform with auto prefix cache
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
All 30 problems from the 2026 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.
PhD-level Google-proof science reasoning questions
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.
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.
OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Kimi K2.6 (1T MoE) Full Benchmark Scorecard
28 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 7 items
Claw-Eval tests real-world agentic task completion across complex multi-step scenarios, evaluating a model's ability to use tools, navigate environments, and complete end-to-end tasks autonomously.
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++.
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.
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 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.
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.
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.
OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.
MCP-Mark evaluates LLMs on their ability to use Model Context Protocol (MCP) tools effectively, testing tool discovery, selection, invocation, and result interpretation across diverse MCP server scenarios.
Tool Decathlon is a comprehensive benchmark for evaluating AI agents' ability to use multiple tools across diverse task categories. It measures proficiency in tool selection, sequencing, and execution across ten different tool-use scenarios.
Finance Agent v2 is an agentic financial-analysis benchmark from Vals that evaluates models on real-world finance workflows, measuring their ability to retrieve and reason over financial documents, perform multi-step calculations, and produce accurate analyses.
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.
📐 Math 6 items
All 30 problems from the 2026 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.
MATH-Vision is a dataset designed to measure multimodal mathematical reasoning capabilities.
HMMT February 2026 is a math competition benchmark based on problems from the Harvard-MIT Mathematics Tournament, testing advanced mathematical problem-solving and reasoning.
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 4 items
A visual reasoning benchmark evaluating multimodal inference under challenging spatial and grounded tasks.
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.
A benchmark for early-stage visual reasoning and perception on child-like vision tasks.
🧠 Reasoning & Science 3 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.
OJBench is a competition-level code benchmark designed to assess the competitive-level code reasoning abilities of large language models. It comprises 232 programming competition problems from NOI and ICPC, categorized into Easy, Medium, and Hard difficulty levels.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ Kimi K2.6 (1T MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 659.9%
Exceeds usable VRAM by 526.75 GB
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run kimi-k2.6:latestInsufficient Local VRAM? Ways to Run Kimi K2.6 (1T MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 Kimi K2.6 (1T MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 2000 GB | 1.15 GB | 2042.45 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 1062.5 GB | 1.15 GB | 1086.2 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 820 GB | 1.15 GB | 838.85 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 711.25 GB | 1.15 GB | 727.93 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 606.25 GB | 1.15 GB | 620.83 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 531.25 GB | 1.15 GB | 544.33 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 488.75 GB | 1.15 GB | 500.98 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 393.75 GB | 1.15 GB | 404.08 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for Kimi K2.6 (1T MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of Kimi K2.6 (1T MoE)?
Kimi K2.6 (1T MoE) uses KimiK25ForConditionalGeneration with 61 layers (61-Layer MLA Multimodal MoE (384 Routed + 1 Shared Expert, 8 Active · 400M ViT)), hidden_size=7168, 64 Q heads / 64 KV heads (head_dim=64), and vocab_size=163,840.
How much VRAM does Kimi K2.6 (1T MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), Kimi K2.6 (1T MoE) requires 620.83 GB VRAM. Full FP16 requires 2042.45 GB.