Kimi K2.6

+ contribute

Kimi K2.6 is an open-weights language model from Moonshot AI, released April 2026. Its published size is 1000B parameters (32B active per token) and a 262k-token context window. Licence: modified-mit.

72.9±0.5#57 of 719 LLMsfrom 56 benchmark results across 17 sources

$ how it compares

The LLMs ranked either side of Kimi K2.6, out of 828 with a score, then more from Moonshot AI.

#	model	score	where	RAM	released
59	Qwen3.8 27B Alibaba	73±1	api $1.60/M +local	~18GB	2026-08
60	Grok 4.20 Multi-Agent (0309) SpaceXAI (xAI)	73±2	api $1.56/M		2026-03
61	Kimi K2.6 Moonshot AI	73±1	api $1.71/M +local	~317GB	2026-04
62	GPT-5 Pro OpenAI	73±1	api $41.25/M		2025-10
63	LongCat 2.0 (1.6T-A48B) Meituan	73±2	api $1.30/M +local	~2155GB	2026-06
	more from Moonshot AI
25	Kimi K3 Moonshot AI	77±1	api $6.00/M +local	~884GB	2026-07
72	Kimi K2.7 Code Moonshot AI	72±1	api $1.71/M +local	~359GB	2026-06
92	Kimi K2.5 Moonshot AI	70±1	api $1.20/M +local	~160GB	2026-01

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #52 of 300 · 20% of its score · 13 tests

Beats Gemini 3.1 Pro (Preview), Qwen3.7 Max, GPT-5.2 here, though they rank higher overall. Trails Claude Sonnet 4.6 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	76.7% ✎	▁··▁·▁▂▁▄█▄▁ beats 22/29	hard · avg 19%	▇▇▇ 6.7%	sharp · several runs
Terminal-Bench 2.1	67.2% ✎	█··▄█·▄····▄ beats 6/6	very hard · avg 12%	▇ 2.0%	at its level
SWE-bench Multilingual	77.1% ✎	▄··▄▄···▄▄·█ beats 6/6	hard · avg 30%	▇ 1.9%	typical
Agents' Last Exam overall full-pass rate (fixed OpenClaw + GUI-as-Tool)	9.2% ✎	▄···█▄▄▄█▄·█ beats 3/10	very hard · avg 1.7%	▇ 1.9%	typical
BrowseComp	61.3% ✎	▄·▄▄···▄··█▄ beats 6/6	hard · avg 16%	▇ 1.2%	at its level
Humanity's Last Exam with tools	54.0% ✎	▄····▄·█·▄·▄ beats 5/5	hard · avg 16%	▇ 1.1%	at its level
Thinkingbox-bench pass@1 (507 stateful business-workflow agent tasks, N=20)	37.7% ✎	▅··▃·▃·█▃▅██ beats 7/15	very hard · avg 4.9%	▇ 1.0%	at its level
τ³-bench	72.4% ✎	▄····▄·█·▄·█ beats 4/6	medium · avg 40%	▇ 0.9%	blunt · at its level

5 more tests

test	result	vs the others in its run	difficulty	counts	why
Legal Research Bench (agentic, all-pass)	15.5% ✎	▄█··▄▄▄██·▄█ beats 2/12	very hard · avg 5.8%	▇ 0.8%	at its level
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)	51.3% ✎	▂·▁▁··▇▇█▂·▁ beats 21/25	hard · avg 29%	▇ 0.7%	blunt · at its level
τ³-bench telecom (tau2-bench ≥ v1.0.1)	97.8% ✎	▄·▄···▄·▄·█▄ beats 4/6	easy · avg 77%	▇ 0.5%	maxed out
KlinikeBench strict pass@1 (333 interactive clinical encounters, Terminus-2)	23.4% ✎	▁▁·▁▁▁▁▅▃█▅█ beats 22/27	very hard · avg 2.4%	▇ 0.5%	typical
TerminalWorld-Verified (200 tasks) with Terminus-2	57.5% ✎	▄··▄·▄·█···▄ beats 4/5	hard · avg 27%	▇ 0.5%	blunt · at its level

coding▁▁▁▃▆██▇▆▄▃▁ #49 of 458 · 17% of its score · 7 tests

Beats GPT-5.4, DeepSeek V4 Flash Vision (Exp), Gemini 3.1 Pro (Preview) here, though they rank higher overall. Trails Mistral Large 4 (public preview), Qwen3.7 Plus, Claude Sonnet 4.6 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified +1	76.7% ✎	▁··▁·▁▂▁▄█▄▁ beats 22/29	hard · avg 19%	▇▇▇ 6.7%	sharp · several runs
Arena WebDev (Code Arena)	1508.7 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 72/119	avg 981	▇▇ 4.0%	sharp
SWE-bench Multilingual	77.1% ✎	▄··▄▄···▄▄·█ beats 6/6	hard · avg 30%	▇ 1.9%	typical
SciCode subproblems	52.0% ✎	▃····▃··█·▃▃ beats 6/6	hard · avg 17%	▇ 1.4%	at its level
LiveCodeBench v6 (window unspecified)	90.2% ✎	██···█··████ beats 4/6	medium · avg 45%	▇ 1.1%	typical
SciCode (level unspecified)	53.5% ✎	▄·····▄█▄▄·▄ beats 6/6	hard · avg 16%	▇ 1.1%	at its level
LiveCodeBench v6 (Feb–May 2025)	86.5% ✎	▄··█····▄▄·█ beats 6/6	medium · avg 45%	▇ 1.0%	typical

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #41 of 393 * · 2.2% of its score · 1 test

Beats GPT-6 Sol, MiMo-V2.6-Flash (309B-A15B), GPT-6 Luna here, though they rank higher overall. Trails MiMo-V2.5-Pro (1T-A42B), GLM-5.1, Grok 4.20 (0309 reasoning) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1461.0 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 331/379	avg 1268	▇ 2.2%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #41 of 398 · 30% of its score · 11 tests

Beats Grok 4.7, GLM-5.3-Flash, Gemini 3.6 Flash here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tiers 1–3 v2	57.2% ✎	▂··▁▁▃▄▄█▅▃▃ beats 42/82	very hard · avg 1.2%	▇▇▇▇ 11%	sharp · at its level
FrontierMath tier 4 v2	25.6% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 27/63	very hard · avg 0.1%	▇▇▇ 6.8%	sharp
MathArena ArXivMath February 2026 (32 problems)	43.0% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 15/24	very hard · avg 9.1%	▇ 2.9%	at its level
MathArena ArXivMath March 2026 (30 problems)	55.8% ✎	█▃▅·█▃▃·▅·▃▃ beats 7/14	very hard · avg 15%	▇ 2.5%	at its level
MathArena ArXivMath January 2026 (23 problems)	71.7% ✎	▂··▄▂▄█▆▆▄▄█ beats 19/23	hard · avg 22%	▇ 1.9%	at its level
MathArena BrokenArXiv March 2026 (56 problems)	15.6% ✎	█·▆█▂·▂▂···▂ beats 9/14	very hard · avg 0.8%	▇ 1.0%	sharp · noisy
MathArena BrokenArXiv February 2026 (31 problems)	11.7% ✎	▃▂▂▃█▃▂·▃··▂ beats 7/16	very hard · avg 0.7%	▇ 1.0%	sharp · noisy
IMO-AnswerBench	91.1% ✎	▄····▄·▄▄·█▄ beats 5/6	medium · avg 40%	▇ 1.0%	typical

3 more tests

test	result	vs the others in its run	difficulty	counts	why
AIME 2026	95.4% ✎	▄·▄·▄▄··█··▄ beats 4/6	medium · avg 56%	▇ 0.7%	sharp · maxed out · noisy
OTIS mock AIME 2024–2025	96.1% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 167/193	very hard · avg 8.5%	▇ 0.7%	sharp · maxed out · noisy
MathConstraint full release, no tools (329 instances)	35.9% ✎	▄··▄▄·▄▄█▄·▄ beats 3/8	very hard · avg 11%	▇ 0.6%	at its level

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #37 of 147 · 3.4% of its score · 4 tests

Beats MiMo-V2.6-Pro, MiMo-V2.6-Flash (309B-A15B), GPT-5.6 Luna here, though they rank higher overall. Trails Qwen3.6 Flash, Qwen3.6 Plus, Qwen3.5 397B-A17B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1264.9 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 114/145	avg 1068	▇ 1.5%	blunt
CharXiv (reasoning questions) lab	80.4% ✎	▂▅▃▅▅▆▆▅▆█▆▂ beats 78/100	medium · avg 39%	▇ 1.1%	small run
Arena document (style controlled)	1449.8 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 11/37	avg 1335	▇ 0.7%	blunt
PureDocBench document parsing (Avg3 over Clean/Digital/Real)	70.1% ✎	▂·▁·▂▂▅█·▁▂▂ beats 15/22	medium · avg 62%	▇ 0.0%	blunt · at its level

web search▂▁▂▆▄█▃▃▁▁·▁ #56 of 80 · 3.1% of its score · 3 tests

Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp	61.3% ✎	▄·▄▄···▄··█▄ beats 6/6	hard · avg 16%	▇ 1.2%	at its level
Humanity's Last Exam with tools	54.0% ✎	▄····▄·█·▄·▄ beats 5/5	hard · avg 16%	▇ 1.1%	at its level
Legal Research Bench (agentic, all-pass)	15.5% ✎	▄█··▄▄▄██·▄█ beats 2/12	very hard · avg 5.8%	▇ 0.8%	at its level

other tests19 not in a task · 36% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond +1	90.8% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 178/208	hard · avg 45%	▇▇ 5.2%	sharp · several runs
Humanity's Last Exam +1	34.8% ✎	▄··█▄··▄·▄·▄ beats 5/6	very hard · avg 5.6%	▇▇ 4.6%	at its level · several runs
EBR-bench (Epoch)	2.4% ✎	▄▂·▆█▂▆▄▆▆·▄ beats 1/23	very hard · avg 0.3%	▇ 3.3%	sharp
Furniture Assembly (Epoch)	21.7% ✎	█▆▆▆▆▂▅▂▂▂·▅ beats 2/30	very hard · avg 1.2%	▇ 3.2%	sharp · at its level
GPQA (subset unspecified)	91.0% ✎	▄█▄▄·▄·····▄ beats 6/6	hard · avg 47%	▇ 3.0%	sharp
Mystery Game Puzzles (Epoch)	18.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 30/74	very hard · avg 9.4%	▇ 2.2%	sharp · noisy
MMLU-Pro	88.1% ✎	▄·····▄▄▄·▄█ beats 5/6	medium · avg 53%	▇ 1.9%	typical
Global-MMLU-Lite	90.6% ✎	▃······▅··█▃ beats 4/6	medium · avg 60%	▇ 1.7%	typical

11 more tests

test	result	vs the others in its run	difficulty	counts	why
Chess Puzzles (Epoch)	26.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 96/135	very hard · avg 5.9%	▇ 1.6%	sharp · noisy
EQ-Bench 3 (emotional intelligence, judge Elo)	1319.2 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 55/62	avg 1044	▇ 1.4%	typical
AgentWorldBench (open-ended, overall)	53.4% ✎	▄▄··▄▄·▄█▄██ beats 4/11	hard · avg 25%	▇ 1.2%	blunt · at its level
MMLU-ProX	85.0% ✎	▄·····▄·▄▄█▄ beats 3/6	medium · avg 57%	▇ 1.1%	typical
IFBench (aggregation unspecified)	76.0% ✎	▃······█▃▃·▃ beats 3/6	medium · avg 44%	▇ 1.1%	typical
IFBench prompt-level (loose)	73.7% ✎	▄▄·▄··▄▄···█ beats 0/6	medium · avg 45%	▇ 1.0%	typical
MultiChallenge	63.1% ✎	▁··········█ beats 2/6	hard · avg 27%	▇ 0.9%	at its level
Legal Research Bench (single-shot, no tools, all-pass)	2.9% ✎	▃▃█▃·▅···█▃▃ beats 2/12	very hard · avg 0.8%	▇ 0.9%	too hard for it
SimpleQA Verified	34.9% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 30/78	very hard · avg 8.0%	▇ 0.9%	at its level
CritPt	9.1% ✎	▄····▄█···█▄ beats 4/6	very hard · avg 1.0%	▇ 0.6%	too hard for it · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	89.2% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 34/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $1.71/M · $0.95 in / $4 out per 1M tokens · 262k context · platform.moonshot.ai ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Kimi K2.6 · needs ~600.5 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Kimi K2.6 · needs ~600.5 GB · from huggingface.co/unsloth
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/unsloth/Kimi-K2.6-GGUF@UD-Q4_K_XL -y
Then run lms ls to see its name, and lms chat followed by that name

MLX (Mac)

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Kimi K2.6 · needs ~657.6 GB, Mac only · from huggingface.co/aidiffuser
chat in the terminal
mlx_lm.chat --model aidiffuser/Kimi-K2.6-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model aidiffuser/Kimi-K2.6-MLX-4bit

Official weights and documentation: huggingface.co/moonshotai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 36%, NVIDIA — Nemotron 3 Ultra tec… 25%, A.X K2 Technical Report (SK T… 11% and 14 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	10	▇▇▇▇▇▇▇▇▇▇▇ 36%	CC BY 4.0
NVIDIA — Nemotron 3 Ultra technical report	17	▇▇▇▇▇▇▇ 25%	CC BY 4.0
A.X K2 Technical Report (SK Telecom), arXiv 2608.30181v1	7	▇▇▇ 11%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	5	▇▇▇ 9%	CC BY-SA 4.0
LMArena leaderboard dataset	4	▇▇▇ 9%	CC BY 4.0
Agents' Last Exam — authors' own fixed-harness model comparison	1	▇ 2%	CC BY 4.0
Legal Research Bench (Vals AI, arXiv:2610.00609v1)	2	▇ 2%	CC-BY-4.0
EQ-Bench 3 canonical leaderboard results	1	▇ 1%	MIT
Qwen/Qwen-AgentWorld-35B-A3B model card	1	▇ 1%	Apache-2.0
Moonshot AI vendor-reported benchmark scores for its own Kimi models (Hugging Face model cards) · lab-reported	1	▇ 1%	Factual vendor-reported score (not licensed content)
One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)	1	▇ 1%	CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)	1	▇ 1%	CC BY 4.0
MathConstraint: Automated Generation of Verified Combinatorial Reasoning Instances for LLMs — authors' own model runs	1	▇ 1%	CC BY 4.0
KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (arXiv:2609.38480v2)	1	▇ 1%	CC BY 4.0
TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks — authors' own Terminus-2 model comparison	1	▇ 0%	CC BY 4.0
How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean, Degraded, and Real-World Settings (arXiv:2605.07492v2)	1	▇ 0%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 72.9