Kimi K2.5

+ contribute

Kimi K2.5 is an open-weights language model from Moonshot AI, released January 2026. Its published size is 1000B parameters (32B active per token) and a 262k-token context window. Licence: modified-mit.

69.8±0.7#85 of 719 LLMsfrom 36 benchmark results across 21 sources

$ how it compares

The LLMs ranked either side of Kimi K2.5, out of 828 with a score, then more from Moonshot AI.

#	model	score	where	RAM	released
90	Inkling Small (266B-A12B) Thinking Machines	70±1	api $0.64/M +local	~86GB	2026-07
91	MiMo-V2-Pro Xiaomi	70±1	comparison only		2026-03
92	Kimi K2.5 Moonshot AI	70±1	api $1.20/M +local	~160GB	2026-01
93	GLM-5 Z.ai	70±1	api $1.55/M +local	~119GB	2026-02
94	Qwen3-Max (2026-01-23) Alibaba	70±1	comparison only		2026-01
	more from Moonshot AI
25	Kimi K3 Moonshot AI	77±1	api $6.00/M +local	~884GB	2026-07
61	Kimi K2.6 Moonshot AI	73±1	api $1.71/M +local	~317GB	2026-04
72	Kimi K2.7 Code Moonshot AI	72±1	api $1.71/M +local	~359GB	2026-06

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #87 of 300 · 31% of its score · 7 tests

Beats Grok 4.20 (0309 reasoning), GPT-5.1, Grok 4.3 here, though they rank higher overall. Trails Hy3 preview (295B-A21B), MiniMax-M2.7, Step 3.5 Flash (196B-A11B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified	73.8% ✎	▁··▁·▁▂▁▄█▄▁ beats 15/29	hard · avg 19%	▇▇▇▇ 10%	sharp · at its level
MCP-Atlas (public set) +1	63.8% ✎	▄······▄▄▄▄█ beats 2/6	hard · avg 25%	▇▇ 6.2%	at its level · several runs
τ²-bench aggregate +1	77.0% ✎	▄····█▄▄···▄ beats 0/5	medium · avg 40%	▇▇ 6.1%	several runs
BrowseComp-ZH	62.3% ✎	▁▁▁▁▁▃·▁▄▅▄█ beats 17/23	hard · avg 18%	▇ 3.0%	at its level
Vibe Code Bench (test split, 50 web apps)	17.5% ✎	▃··▃·▃██▃▃▃█ beats 6/14	very hard · avg 2.3%	▇ 2.2%	sharp · at its level · noisy
SWE Atlas (Codebase Q&A, Test Writing, Refactoring; mini-SWE-agent) pass@1	19.1% ✎	▄·▄··█···▄·█ beats 1/6	very hard · avg 6.1%	▇ 2.1%	typical
τ²-bench Telecom	99.7% ✎	▁·▅█▆▅▂▃▃▂▃▃ beats 89/89	hard · avg 16%	▇ 1.3%	sharp · noisy · small run

coding▁▁▁▃▆██▇▆▄▃▁ #73 of 458 · 22% of its score · 4 tests

Beats GPT-5.2, GPT-5, Grok 4.20 (0309 reasoning) here, though they rank higher overall. Trails Gemini 3.5 Flash-Lite, MiMo-V2.5 (310B-A15B), MiMo-V2.5-Pro (1T-A42B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified	73.8% ✎	▁··▁·▁▂▁▄█▄▁ beats 15/29	hard · avg 19%	▇▇▇▇ 10%	sharp · at its level
Arena WebDev (Code Arena)	1435.8 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 55/119	avg 981	▇▇▇ 8.0%	sharp
Vibe Code Bench (test split, 50 web apps)	17.5% ✎	▃··▃·▃██▃▃▃█ beats 6/14	very hard · avg 2.3%	▇ 2.2%	sharp · at its level · noisy
SWE Atlas (Codebase Q&A, Test Writing, Refactoring; mini-SWE-agent) pass@1	19.1% ✎	▄·▄··█···▄·█ beats 1/6	very hard · avg 6.1%	▇ 2.1%	typical

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #55 of 393 * · 4.5% of its score · 1 test

Beats GPT-6 Luna, Grok 4.7, GPT-5.2 here, though they rank higher overall. Trails Gemma 4 31B, Claude Sonnet 4.5, Gemini 3.5 Flash-Lite here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1450.1 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 315/379	avg 1268	▇▇ 4.5%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #69 of 398 · 12% of its score · 5 tests

Beats Claude Sonnet 4.6, Claude Opus 4.5, Qwen3.7 Plus here, though they rank higher overall. Trails INTELLECT-3, EXAONE 4.5 33B, Qwen3 235B-A22B Thinking 2507 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MathArena ArXivMath January 2026 (23 problems)	62.5% ✎	▂··▄▂▄█▆▆▄▄█ beats 15/23	hard · avg 22%	▇▇ 4.1%	at its level
MathArena ArXivMath December 2025 (17 problems)	41.9% ✎	▃·▃▃█▁·▄·▁▁▁ beats 11/18	hard · avg 19%	▇ 2.7%	at its level
OTIS mock AIME 2024–2025	92.2% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 151/193	very hard · avg 8.5%	▇ 2.2%	sharp · noisy
MATH-Vision	84.2% ✎	█·····▄▄·▄·▄ beats 3/5	medium · avg 46%	▇ 1.9%	typical
MathArena BrokenArXiv February 2026 (31 problems)	2.8% ✎	▃▂▂▃█▃▂·▃··▂ beats 0/16	very hard · avg 0.7%	▇ 1.4%	sharp · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #42 of 147 · 14% of its score · 7 tests

Beats MiMo-V2.6-Pro, MiMo-V2.6-Flash (309B-A15B), DeepSeek V4 Flash Vision (Exp) here, though they rank higher overall. Trails Qwen3.6 Flash, Gemini 3.5 Flash-Lite here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1252.2 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 105/145	avg 1068	▇ 3.1%	blunt
ChartArena chart parsing (English, mAP_high average)	54.8% ✎	▁·▁·▃▃▁▇▃█·▄ beats 21/22	hard · avg 27%	▇ 2.5%	blunt · at its level
CharXiv (reasoning questions) lab	77.5% ✎	▂▅▃▅▅▆▆▅▆█▆▂ beats 68/100	medium · avg 39%	▇ 2.3%	small run
MATH-Vision	84.2% ✎	█·····▄▄·▄·▄ beats 3/5	medium · avg 46%	▇ 1.9%	typical
SciDocBench scientific-document workflows (overall, 496 instances)	50.4% ✎	▃▃···█▃▅▃·▃▅ beats 3/11	hard · avg 26%	▇ 1.5%	blunt · at its level
Arena document (style controlled)	1436.0 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 6/37	avg 1335	▇ 1.5%	blunt
Diagram-MMU diagram question answering (DQA accuracy, scientific TikZ diagrams)	79.7% ✎	▄·····▄▄·█·█ beats 1/6	medium · avg 55%	▇ 0.8%	blunt

web search██···█····██ #3 of 5 * · 3.0% of its score · 1 test

Trails Kimi K2 Thinking, Step 3.5 Flash (196B-A11B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp-ZH	62.3% ✎	▁▁▁▁▁▃·▁▄▅▄█ beats 17/23	hard · avg 18%	▇ 3.0%	at its level

other tests14 not in a task · 33% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond	87.6% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 157/208	hard · avg 40%	▇▇ 5.6%	sharp
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)	12.1% ✎	█··▂▅▃▃▁▃·▁▃ beats 13/24	very hard · avg 0.3%	▇ 3.6%	sharp · noisy
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	73.1% ✎	▁·▇▃▁·█▇▁·▇▃ beats 16/27	very hard · avg 4.1%	▇ 3.4%	sharp · at its level · noisy
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	40.2% ✎	▁▁▂▁▃▄█▆▄▅▂▃ beats 34/39	very hard · avg 22%	▇ 3.2%	blunt · at its level
DSEffi-Bench Pass (1,000 data-science tasks, greedy)	40.2% ✎	▅··▃▅·█·▃██▃ beats 5/15	hard · avg 24%	▇ 3.2%	at its level
EQ-Bench 3 (emotional intelligence, judge Elo)	1303.5 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 46/62	avg 1044	▇ 2.7%	typical
WildClawBench	30.8% ✎	▃·▅▅█·▅█▅▃▃▃ beats 4/17	very hard · avg 11%	▇ 2.5%	at its level
Chess Puzzles (Epoch)	12.0% ✎	█▁▁▁▁▃▁▄▂▂▁▁ beats 58/135	very hard · avg 5.9%	▇ 2.4%	sharp · noisy

6 more tests

test	result	vs the others in its run	difficulty	counts	why
ResearchRubrics	59.5% ✎	███··██·█··█ beats 3/6	hard · avg 31%	▇ 2.0%	at its level
SimpleQA Verified	34.3% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 29/78	very hard · avg 8.0%	▇ 1.7%	at its level
GAIA (no-file subset)	75.9% ✎	▃▃····█▃···▃ beats 4/6	medium · avg 47%	▇ 1.0%	typical
SOOHAK Challenge (340 problems, Avg@3)	13.9% ✎	█·█████··█·█ beats 5/7	very hard · avg 4.1%	▇ 0.8%	typical
SOOHAK-Mini (702 problems, Avg@3)	66.1% ✎	▄··▄·▄·▄▄█·▄ beats 5/7	medium · avg 36%	▇ 0.6%	blunt · at its level
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	85.8% ✎	▁▂▁▃▅█▂▃▄▃▁▁ beats 14/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · Amazon Bedrock · $1.20/M · $0.6 in / $3 out per 1M tokens · 262k context · docs.aws.amazon.com ↗ · price sourceUSD per 1M tokens, Amazon Bedrock on-demand Standard tier, Kimi K2.5 (Moonshot AI section, 'Kimi 2' table), US East (N. Virginia), US East (Ohio) and US West (Oregon). Priority (+75%) and Flex (-50%) tiers, cache and other regions excluded; Kimi K2.5 is not on Bedrock's model-lifecycle (legacy/EOL) page. Moonshot's own Kimi platform no longer lists kimi-k2.5. Currency: USD. Amazon Bedrock official pricing page, observed 2026-09-25.

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 K2.5 · needs ~600.5 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/Kimi-K2.5-GGUF:UD-Q4_K_XL
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Kimi-K2.5-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 K2.5 · 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-K2.5-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 K2.5 · needs ~657.6 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/Kimi-K2.5
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/Kimi-K2.5

Official weights and documentation: huggingface.co/moonshotai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 21%, LMArena leaderboard dataset 17%, stepfun-ai/Step-3.5-Flash mod… 9% and 18 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	5	▇▇▇▇▇▇ 21%	CC BY 4.0
LMArena leaderboard dataset	4	▇▇▇▇▇ 17%	CC BY 4.0
stepfun-ai/Step-3.5-Flash model card	4	▇▇▇ 9%	Apache-2.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	3	▇▇ 8%	CC BY-SA 4.0
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	2	▇▇ 7%	MIT
Qwen/Qwen3.5-397B-A17B model card	2	▇▇ 5%	apache-2.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 3%	CC BY 4.0
DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation — authors' own model runs	1	▇ 3%	CC BY 4.0
zai-org/GLM-5 model card	1	▇ 3%	MIT
GLM-5 Team (Zhipu AI) — GLM-5 technical report	1	▇ 3%	CC BY 4.0
EQ-Bench 3 canonical leaderboard results	1	▇ 3%	MIT
ChartArena: Benchmarking Chart Parsing across Languages, Scenarios, and Formats (arXiv:2606.01348v3)	1	▇ 2%	CC BY 4.0
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation (arXiv:2605.10912v1)	1	▇ 2%	CC BY 4.0
Moonshot AI vendor-reported benchmark scores for its own Kimi models (Hugging Face model cards) · lab-reported	1	▇ 2%	Factual vendor-reported score (not licensed content)
Vibe Code Bench (Vals AI, arXiv:2603.04601v3)	1	▇ 2%	CC-BY-4.0
SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution (arXiv:2605.08366v1)	1	▇ 2%	CC BY 4.0
SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding (arXiv:2609.05141v2)	1	▇ 2%	CC BY 4.0
Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs — authors' own model runs	2	▇ 1%	CC BY-SA 4.0
SEATauBench: Progressively Adapting Tool-Agent-User Evaluation Into Low-Resource Southeast Asian Languages (arXiv:2606.28715v2)	1	▇ 1%	CC-BY-4.0
Diagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams (arXiv:2608.12262v1)	1	▇ 1%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 69.8