Parameters
1.0T total / 32.0B active
MoE: total / active
Architecture
Mixture-of-Experts (MoE)
Released
14.04.2026
License
Modified MIT License (Kimi)
Input Modalities
Output Modalities
Context (native)
262,144 tokens
Context (extended)
262,144 tokens
About
Kimi K2.6 (moonshotai/Kimi-K2.6) is Moonshot AI's open-source native multimodal agentic model - a trillion-parameter Mixture-of-Experts with 32B activated parameters that advances long-horizon coding, coding-driven design, proactive autonomous execution and swarm-based task orchestration. It accepts text, image and video input, ships with native INT4 quantization (same method as Kimi-K2-Thinking), and supports both thinking mode and instant mode.
The architecture matches the Kimi K2.5 line: 61 layers (1 dense), attention hidden dim 7168 with Multi-head Latent Attention (MLA) (64 heads), 384 experts with 8 selected per token plus 1 shared expert (MoE expert dim 2048, SwiGLU), 160K vocabulary, 256K context, and the 400M MoonViT vision encoder.
Key features: Long-Horizon Coding with significant gains on complex end-to-end coding tasks across Rust, Go and Python (front-end, DevOps, performance optimization); Coding-Driven Design turning simple prompts and visual inputs into production-ready interfaces and full-stack workflows with structured layouts and rich animations; an elevated Agent Swarm scaling horizontally to 300 sub-agents executing 4,000 coordinated steps - dynamically decomposing tasks into parallel, domain-specialized subtasks; and Proactive & Open Orchestration powering persistent 24/7 background agents that manage schedules, execute code and orchestrate cross-platform operations without human oversight. Released April 14, 2026.
Training Data Native multimodal agentic model. Native INT4 quantization (same method as Kimi-K2-Thinking). Supports thinking mode and instant mode. Agent swarm capability scaling to 300 sub-agents executing 4,000 coordinated steps.
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
Terminal Bench 2.1
coding_agent
|
59.31%
|
11.06.2026 |
#30/44
better than 33%
|
|
NL2Repo-Bench
coding_agent
|
39.48%
|
11.06.2026 |
#16/23
better than 32%
|
|
VIBE-V2
coding_agent
|
64.52%
|
11.06.2026 |
#5/9
better than 50%
|
|
SVG-Bench
coding_agent
|
74.53%
|
11.06.2026 |
#5/9
better than 50%
|
|
SpreadSheetBench-v1
general_agent
|
91.03%
|
11.06.2026 |
#7/9
better than 25%
|
|
Claw-Eval Avg
coding_agent
|
73.78%
|
11.06.2026 |
#18/40
better than 56%
|
|
OSWorld-Verified
general_agent
|
76.22%
|
11.06.2026 |
#10/25
better than 62%
|
|
AgentWorldBench MCP
general_agent
|
67.90%
|
24.06.2026 |
#7/14
better than 54%
|
|
AgentWorldBench Search
general_agent
|
34.89%
|
24.06.2026 |
#10/14
better than 31%
|
|
AgentWorldBench Terminal
general_agent
|
62.19%
|
24.06.2026 |
#8/14
better than 46%
|
|
AgentWorldBench SWE
general_agent
|
68.70%
|
24.06.2026 |
#11/14
better than 23%
|
|
AgentWorldBench Android
general_agent
|
69.91%
|
24.06.2026 |
#7/14
better than 54%
|
|
AgentWorldBench Web
general_agent
|
41.01%
|
24.06.2026 |
#11/14
better than 23%
|
|
AgentWorldBench OS
general_agent
|
32.52%
|
24.06.2026 |
#10/14
better than 31%
|
|
AgentWorldBench Overall
general_agent
|
57.98%
|
24.06.2026 |
#9/14
better than 38%
|
|
Vals.ai Financial Agent 1.1 (without web search)
general_agent
|
27.00%
|
04.06.2026 |
#6/7
better than 17%
|
|
Terminal-Bench 2.1 (Terminus-2)
coding_agent
|
68.14%
|
04.06.2026 |
#21/44
better than 53%
|
|
Vals.ai Financial Agent 1.1 (with web search)
general_agent
|
70.34%
|
04.06.2026 |
#5/7
better than 33%
|
|
CritPt (no tools)
stem_reasoning
|
26.81%
|
04.06.2026 |
#15/21
better than 30%
|
|
MMLU-Pro
knowledge
|
90.65%
|
04.06.2026 |
#4/48
better than 94%
|
|
IOI 2025
stem_reasoning
|
100.00%
|
04.06.2026 |
#1/5
better than 100%
|
|
LiveCodeBench v6
stem_reasoning
|
95.50%
|
04.06.2026 |
#8/49
better than 85%
|
|
SWE-bench Verified
coding_agent
|
86.35%
|
04.06.2026 |
#22/60
better than 64%
|
|
SWE-bench Multilingual
coding_agent
|
85.33%
|
04.06.2026 |
#12/48
better than 77%
|
|
IMOAnswerBench
stem_reasoning
|
97.30%
|
04.06.2026 |
#3/30
better than 93%
|
|
ProfBench (Search)
general_agent
|
71.94%
|
04.06.2026 |
#3/7
better than 50%
|
|
IMOAnswerBench (with tools)
stem_reasoning
|
100.00%
|
04.06.2026 |
#1/7
better than 100%
|
|
Apex-Shortlist (no tools)
stem_reasoning
|
79.12%
|
04.06.2026 |
#7/11
better than 40%
|
|
IFBench (prompt loose)
instruction_following
|
57.44%
|
04.06.2026 |
#12/14
better than 15%
|
|
PinchBench
general_agent
|
91.97%
|
04.06.2026 |
#2/9
better than 88%
|
|
TauBench V3 Airline
general_agent
|
100.00%
|
04.06.2026 |
#1/7
better than 100%
|
|
Apex-Shortlist (with tools)
stem_reasoning
|
61.56%
|
04.06.2026 |
#5/7
better than 33%
|
|
Multi-Challenge
instruction_following
|
95.37%
|
04.06.2026 |
#7/18
better than 65%
|
|
TauBench V3 Retail
general_agent
|
82.90
|
04.06.2026 |
#7/7
better than 0%
|
|
TauBench V3 Telecom
general_agent
|
94.25%
|
04.06.2026 |
#3/7
better than 67%
|
|
GPQA Diamond
stem_reasoning
|
93.76%
|
04.06.2026 |
#22/80
better than 73%
|
|
AA-LCR
long_context
|
83.57%
|
04.06.2026 |
#10/42
better than 78%
|
|
SciCode (subtask)
stem_reasoning
|
66.50%
|
04.06.2026 |
#4/15
better than 79%
|
|
Humanity's Last Exam
stem_reasoning
|
61.93%
|
04.06.2026 |
#28/71
better than 60%
|
|
MMLU-ProX
multilingual
|
90.97%
|
04.06.2026 |
#4/19
better than 83%
|
|
TauBench V3 Average
general_agent
|
82.89%
|
04.06.2026 |
#3/7
better than 67%
|
|
MMMU-Pro
vision_language
|
91.85%
|
11.06.2026 |
#7/35
better than 82%
|
|
HLE (with tools)
stem_reasoning
|
71.48%
|
04.06.2026 |
#19/49
better than 62%
|
|
Kimi Code Bench V2
coding_agent
|
50.90
|
23.08.2026 |
#7/7
better than 0%
|
|
MLS-Bench-Lite
coding_agent
|
26.70
|
23.08.2026 |
#9/9
better than 0%
|
|
Kimi Claw 24/7 Bench
agentic
|
42.90
|
23.08.2026 |
#4/4
better than 0%
|
|
MCPMark-Verified
agentic
|
72.80
|
23.08.2026 |
#4/4
better than 0%
|
|
BrowseComp Agent Swarm
general_agent
|
100.00%
|
23.08.2026 |
#1/2
better than 100%
|
|
WideSearch
general_agent
|
89.24%
|
23.08.2026 |
#10/24
better than 61%
|
|
Toolathlon Verified
general_agent
|
44.31%
|
23.08.2026 |
#23/35
better than 35%
|
|
MCPMark
general_agent
|
58.13%
|
23.08.2026 |
#4/15
better than 71%
|
|
Apex-Agents
general_agent
|
34.74%
|
23.08.2026 |
#18/24
better than 26%
|
|
Terminal-Bench 2.0
coding_agent
|
95.08%
|
23.08.2026 |
#3/23
better than 91%
|
|
AIME 26
stem_reasoning
|
96.43%
|
23.08.2026 |
#11/45
better than 77%
|
|
HMMT Feb 26
stem_reasoning
|
93.52%
|
23.08.2026 |
#13/44
better than 72%
|
|
CharXiv (RQ)
document_understanding
|
56.92%
|
23.08.2026 |
#11/34
better than 70%
|
|
MathVision
vision_language
|
79.92%
|
23.08.2026 |
#7/27
better than 77%
|
|
BabyVision
vision_language
|
33.37%
|
23.08.2026 |
#14/24
better than 43%
|
|
V-Star
vision_language
|
96.23%
|
23.08.2026 |
#2/20
better than 89%
|
|
GDPVal-AA v2
general_agent
|
80.94%
|
26.06.2026 |
#21/38
better than 46%
|
|
Program Bench
coding_agent
|
58.48%
|
23.08.2026 |
#4/7
better than 50%
|
|
BrowseComp
general_agent
|
65.39%
|
04.06.2026 |
#22/45
better than 52%
|
|
OJBench
stem_reasoning
|
85.17%
|
23.08.2026 |
#2/22
better than 95%
|
|
DeepSearch QA
general_agent
|
92.49%
|
23.08.2026 |
#2/9
better than 88%
|
|
WMT24++ (en→xx)
multilingual
|
95.41%
|
04.06.2026 |
#4/17
better than 81%
|
|
TauBench V3 Banking
general_agent
|
50.10%
|
04.06.2026 |
#9/17
better than 50%
|
|
MCP-Atlas
general_agent
|
70.50%
|
11.06.2026 |
#33/54
better than 40%
|
|
GDPVal
general_agent
|
31.47%
|
04.06.2026 |
#16/20
better than 21%
|
|
OmniScience Accuracy
knowledge
|
71.56%
|
04.06.2026 |
#4/10
better than 67%
|
|
OmniScience Non-Hallucination
knowledge
|
71.34%
|
04.06.2026 |
#4/10
better than 67%
|
|
SWE-bench Pro
coding_agent
|
65.33%
|
11.06.2026 |
#20/77
better than 72%
|
|
Claw-Eval Pass^3
coding_agent
|
69.98%
|
23.08.2026 |
#4/21
better than 85%
|
What is a good Terminal Bench 2.1 score? What is a good NL2Repo-Bench score? What is a good VIBE-V2 score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 32.0B are active per token
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 2000.0 GB | 5.0 GB | 2105.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 1000.0 GB | 2.5 GB | 1052.5 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 500.0 GB | 1.2 GB | 526.2 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 500.0 GB | 1.2 GB | 526.2 GB | Which GPUs can run it? Calculate on your GPU |
Only datacenter GPUs — FP16 needs 2105.0 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.
Architecture
- Attention
- Multi-head Latent Attention
- MoE
- 384 experts · top-8 per token
- Layers
- 61
- Hidden size
- 7168
- Context
- 262K tokens
- RoPE θ
- 50K
- Parameters
- 1000000M
- Active params
- 32000M
Source: Hugging Face config.json · KimiK25ForConditionalGeneration · model repo
Training Pipeline
-
1
pretraining
Native multimodal agentic pre-training (Kimi K2.6 line)
K2.6 continues the Kimi K2 multimodal agentic lineage (K2.5 was built via ~15T-token continual pretraining); card documents capabilities, not token counts.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| Agentic multimodal training data (long-horizon coding, design, swarm) | finetune | — | — |
Linked Resources
Kimi K2.6 HuggingFace Model Card
https://huggingface.co/moonshotai/Kimi-K2.6
Kimi K2.6 Tech Blog
https://www.kimi.com/blog/kimi-k2-6.html
Kimi K2.6 Chat Demo
https://www.kimi.com
Moonshot AI Homepage
https://www.moonshot.ai
Moonshot AI API Platform
https://platform.moonshot.ai
Kimi Code CLI - Agent Framework
https://www.kimi.com/code
Kimi Vendor Verifier
https://kimi.com/blog/kimi-vendor-verifier.html
Kimi K2.5 Technical Report
https://arxiv.org/pdf/2602.02276
Kimi K2-Thinking - Native INT4 Quantization Reference
https://huggingface.co/moonshotai/Kimi-K2-Thinking#4-native-int4-quantization
Kimi K2.6 HuggingChat Demo
https://huggingface.co/spaces/akhaliq/Kimi-K2.6
Claude Opus 4.7 System Card (DeepSearchQA scores source)
https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf
IMO-AnswerBench scores source (z.ai/blog/glm-5.1)
https://z.ai/blog/glm-5.1
APEX-Agents Evaluation by Artificial Analysis
https://artificialanalysis.ai/evaluations/apex-agents-aa
Kimi K2.6 Model Deployment Guide
https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/docs/deploy_guidance.md
Kimi K2.6 License (Modified MIT)
https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/LICENSE
Kimi K2 Thinking Documentation (Interleaved Thinking & Multi-Step Tool Call)
https://platform.moonshot.ai/docs/guide/use-kimi-k2-thinking-model#complete-example
Trend Analysis
24h Change
+0.2%
7d Change
+2.0%
30d Change
+7.3%
Current
20,989
downloads
-2.1%
downloads
+0.2%
likes
+0.1%
downloads_all_time
+0.0%
No comments yet
Be the first to share your thoughts!