whatmodel.sh / K-EXAONE (236B-A23B)

K-EXAONE (236B-A23B)

+ contribute

K-EXAONE (236B-A23B) is an open-weights language model from LG AI Research, released December 2025. Its published size is 236B parameters (23B active per token) and a 262k-token context window. Licence: k-exaone.

Catalogue identity notes

LG AI Research's first K-EXAONE release, repo LGAI-EXAONE/K-EXAONE-236B-A23B @61e6d57 (created 2025-12-26, main history from 2025-12-31; licence tag other / k-exaone). Printed as 'K-EXAONE (Reasoning)' (236B / 23B active) in its own card and as 'K-EXAONE' (236B/23B) in the K-EXAONE 2.0 card. FriendliAI is linked from the card as an API partner but no first-party API page was verified.

Model information sources

Release date ↗ · Weights licence ↗

Model card @61e6d57: '236B in total and 23B activated' (safetensors metadata 237,099,669,632 incl. MTP).

65.3±0.9#141 of 719 LLMsfrom 25 benchmark results across 3 sources

$ how it compares

The LLMs ranked either side of K-EXAONE (236B-A23B), out of 828 with a score, then more from LG AI Research.

#	model	score	where	RAM	released
160	Claude Sonnet 4 Anthropic	65±1	comparison only		2025-05
161	MiMo-V2.6-Distill-Qwen-9B Xiaomi	65±2	local	~7.8GB	2026-09
162	K-EXAONE (236B-A23B) LG AI Research	65±1	local	~152GB	2025-12
163	DeepSeek R1 0528 DeepSeek	65±1	api $0.91/M +local	~110GB	2025-05
164	EXAONE 4.5 33B LG AI Research	65±2	local	~24GB	2026-04
	more from LG AI Research
127	K-EXAONE 2.0 (750B-A37B) LG AI Research	67±1	local	~791GB	2026-07
251	EXAONE 4.0 32B LG AI Research	59±1	local	~20GB	2025-07
296	EXAONE Deep 32B LG AI Research	56±2	local	~22GB	2025-03

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #122 of 300 · 24% of its score · 7 tests

Beats o3, Gemini 2.5 Pro, Gemini 3.1 Flash-Lite here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Large, Nova 2 Pro (Preview) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified lab	49.4% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 40/131	hard · avg 20%	▇▇ 5.5%	sharp · at its level · small run
τ²-bench Telecom lab	73.5% ✎	▁·▅█▆▅▃▃▃▂▃▁ beats 58/89	hard · avg 17%	▇▇ 4.7%	sharp · at its level · small run
Terminal-Bench 2.1 lab	30.3% ✎	▁·▁·▁▁▄▃▄█▃▇ beats 9/56	very hard · avg 11%	▇▇ 4.1%	at its level · small run
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	10.3% ✎	▃····█··▃█·▃ beats 1/8	very hard · avg 3.4%	▇▇ 3.9%	typical
τ²-bench Retail lab	78.6% ✎	▁····▂▁▃▆█▃▁ beats 40/44	medium · avg 44%	▇ 2.7%	at its level · small run
τ²-bench Airline lab	60.4% ✎	▁▁·▂▁▃▂▁█▅▂▃ beats 32/42	hard · avg 31%	▇ 2.3%	at its level · small run
τ³-bench Banking lab	14.2% ✎	▂▂·▅█▂▂▃▂▅▂▃ beats 12/20	very hard · avg 5.6%	▇ 1.2%	blunt · small run

coding▁▁▁▃▆██▇▆▄▃▁ #125 of 458 · 14% of its score · 3 tests

Beats Qwen3.5 Plus, Gemini 2.5 Pro, Gemini 3.1 Flash-Lite here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Seed-OSS 36B Instruct, Gemma 4 26B-A4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Verified lab	49.4% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 40/131	hard · avg 20%	▇▇ 5.5%	sharp · at its level · small run
LiveCodeBench Pro 2025Q2 (Medium) lab	25.9% ✎	█··▄▂▂██▄·▄▆ beats 21/22	very hard · avg 3.0%	▇▇ 4.6%	sharp · small run
LiveCodeBench v6 (window unspecified) lab	80.7% ✎	▂▃▄▄▅██▅▄▇▃▆ beats 45/65	medium · avg 46%	▇ 3.4%	small run

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #132 of 393 * · 3.9% of its score · 1 test

Beats o3-mini here, though they rank higher overall. Trails LFM2.5 8B-A1B, LFM2.5 2.6B, Nemotron-Cascade-8B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
IFEval (aggregation unspecified) lab +1	89.7% ✎	▄·▄·▄·▄▄···█ beats 6/6	easy · avg 80%	▇▇ 3.9%	several runs

math▁▂▃▅▆▅▄█▄▁▁▁ #97 of 398 · 31% of its score · 5 tests

Beats Claude Sonnet 4.6, MiniMax-M3, Nemotron 3 Ultra 550B-A55B here, though they rank higher overall. Trails Gemini 2.5 Flash, GLM-4.5-Air, Granite 4.2 30B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2026 lab	92.2% ✎	▁▁▁▂▁▂▆█▆▂·▁ beats 28/49	medium · avg 60%	▇▇▇ 8.1%	sharp · small run
IMO-AnswerBench lab +1	76.3% ✎	▁▁·▁▁▁▄▅█▄▂▃ beats 15/42	medium · avg 38%	▇▇▇ 8.1%	small run · several runs
HMMT February 2026 lab	80.7% ✎	▃▃▂▃·██▆█▅▃▅ beats 19/33	hard · avg 34%	▇▇▇ 7.4%	sharp · small run
AIME 2025 lab	92.8% ✎	█·▄▄····▄··█ beats 5/6	medium · avg 53%	▇▇ 4.9%	sharp
HMMT November 2025 lab	86.8% ✎	▁▁····▁▅▃█▇▁ beats 8/23	easy · avg 65%	▇ 2.2%	small run

web search▂▃▂▆▄█▃▃▁▁·▁ #76 of 80 * · 3.9% of its score · 1 test

Trails GPT-4.1 nano, GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	10.3% ✎	▃····█··▃█·▃ beats 1/8	very hard · avg 3.4%	▇▇ 3.9%	typical

other tests4 not in a task · 33% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond lab +1	79.1% ✎	█···▄▄···▄·█ beats 5/6	hard · avg 51%	▇▇▇▇▇▇ 14%	sharp · several runs
MMLU-Pro lab +1	83.8% ✎	▁▁·▁▁▁▃▂▄▅█▁ beats 159/188	medium · avg 56%	▇▇▇ 7.7%	small run · several runs
Humanity's Last Exam lab +1	13.6% ✎	▁▁▁▃▆▆▇██▆▁▃ beats 30/104	very hard · avg 4.4%	▇▇▇ 6.9%	small run · several runs
IFBench (aggregation unspecified) lab +1	67.3% ✎	▃▇▂▄▅▅▂▄▄▆█▇ beats 51/80	medium · avg 41%	▇▇ 3.8%	at its level · small run · several runs

$ where to run it

llama.cpp · recommended

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 K-EXAONE (236B-A23B) · needs ~143.47 GB · from huggingface.co/LGAI-EXAONE
chat in the terminal
llama cli -hf LGAI-EXAONE/K-EXAONE-236B-A23B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf LGAI-EXAONE/K-EXAONE-236B-A23B-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run K-EXAONE (236B-A23B) · needs ~143.47 GB · from huggingface.co/LGAI-EXAONE
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B-GGUF@Q4_K_M -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 K-EXAONE (236B-A23B) · needs ~251.4 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/K-EXAONE-236B-A23B-8bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/K-EXAONE-236B-A23B-8bit

Other ways to run it

Run it as published · needs ~474.2 GB, NVIDIA only · from huggingface.co/LGAI-EXAONE
vllm serve LGAI-EXAONE/K-EXAONE-236B-A23B --reasoning-parser deepseek_v3 --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser hermes

Weights and documentation: huggingface.co/LGAI-EXAONE ↗

$ how the score was worked out · LGAI-EXAONE/K-EXAONE-2.0-750B… 50%, LG AI Research vendor-reporte… 46%, K-BrowseComp: A Web Browsing… 4%

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
LGAI-EXAONE/K-EXAONE-2.0-750B-A37B model card · lab-reported	11	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 50%	Apache-2.0
LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards) · lab-reported	13	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 46%	Factual vendor-reported score (not licensed content)
K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts (arXiv:2606.02404v1)	1	▇ 4%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 65.3

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json