The story so far
Weights, every dated article, and repo commits on one rail.
18 repo commits since the weights landed
Teal is the weights (2025-12-26), grey ticks are repo commits.
K-EXAONE-236B-A23B
LGAI-EXAONE k-exaone out 10 months ago
237B parameters, 23B active (8 of 128 experts per token), text in, text out, 256K tokens of context.
Verified 2026-10-25 Rank #747 in LLMs · score 71.5 method v1 Sources 3 Downloads 30d 11.7k · #500 of 619 tracked
Can I use it?k-exaoneCustom license; read the repo LICENSE before commercial hostinglicence file2026-10-24
Do people use it?11.7kdownloads, 30 days#500 of 619 we track · below the median577 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 13 quantized builds and 5 Spaces.
- As a service0no API provider listed
- On own hardware13quantized builds
- Built on31 fine-tunes · 2 adapters · 0 merges
- In apps5Hugging Face Spaces
Top Spaces running it: mayafree/Model-Genome-Korea ♥ 33 · hfmlsoc/different-flops ♥ 11
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits two 80 GB GPUs at INT4; too big for one
About 142 GB at INT4, 285 GB at FP8, 569 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
16 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
16 derivatives on the Hub, most of them quantized
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
18 rows
Benchmarks 18 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (18)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported AA-LCR · Read long documents | 53.5 | Reasoning | model author | 2026-10-07 |
| author-reported AIME 2025 · Reason through hard problems | 92.8 | Reasoning | model author | 2026-10-07 |
| author-reported BrowseComp | 31.4 | Reasoning | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 79.1 | Reasoning | model author | 2026-10-07 |
| author-reported HMMT · Reason through hard problems | 86.8 | Reasoning | model author | 2026-10-07 |
| author-reported Humanity's Last Exam · Reason through hard problems | 13.6 | Reasoning | model author | 2026-10-07 |
| author-reported IFBench · Chat and follow instructions | 67.3 | Reasoning | model author | 2026-10-07 |
| author-reported IFEval · Chat and follow instructions | 89.7 | Reasoning | model author | 2026-10-07 |
| author-reported IMO-AnswerBench · Reason through hard problems | 76.3 | Reasoning | model author | 2026-10-07 |
| author-reported LiveCodeBench · Write code with agents | 25.9 | Reasoning | model author | 2026-10-07 |
| author-reported LiveCodeBench · Write code with agents | 80.7 | Reasoning | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 83.8 | Reasoning | model author | 2026-10-07 |
| author-reported MRCR · Read long documents | 52.3 | Reasoning | model author | 2026-10-07 |
| author-reported SWE-bench Verified · Write code with agents | 49.4 | Reasoning | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 29 | Reasoning | model author | 2026-10-07 |
| author-reported τ²-Bench (Airline) | 60.4 | Reasoning | model author | 2026-10-07 |
| author-reported τ²-Bench (Retail) | 78.6 | Reasoning | model author | 2026-10-07 |
| author-reported τ²-Bench (Telecom) | 73.5 | Reasoning | model author | 2026-10-07 |
What the card says
1 cited sentence
Use-case evidence
5 use cases
What it is good at 5 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Reason through hard problemsauthor-reportedAuthor reports MMLU-Pro 83.8Author reports GPQA-Diamond 79.1Author reports Humanity's Last Exam 13.6+4 more
- Chat and follow instructionsauthor-reportedAuthor reports IFBench 67.3Author reports IFEval 89.7decision-model estimate p=0.88+2 more
- Read long documentsauthor-reportedAuthor reports AA-LCR 53.5Author reports OpenAI-MRCR 52.3decision-model estimate p=1.00+1 more
- Write code with agentsauthor-reportedAuthor reports LiveCodeBench v6 80.7Author reports Terminal-Bench 2.0 29.0Author reports SWE-Bench Verified 49.4
- Call tools and APIsstructuraldecision-model estimate p=0.96chat template: <tool_call>
Facts
25 fields
Facts
- K-EXAONE-236B-A23B has 237B parameters (23B active per token). (safetensors metadata, observed 2026-10-24)
- K-EXAONE-236B-A23B is a mixture-of-experts model with 128 experts, routing 8 per token. (config.json, observed 2026-10-24)
- Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
- License: k-exaone (friction: review). Custom license; read the repo LICENSE before commercial hosting. (license file, observed 2026-10-24)
- Created 2025-12-26, last modified 2026-08-03. (HF API, observed 2026-10-24)
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | ExaoneMoeForCausalLM |
| Model type | exaone_moe |
| Dtype | bfloat16 |
| Params by dtype | BF16: 237,099,663,616 · F32: 6,016 |
| Size band | XXL |
| MoE | yes |
| Experts | 128 |
| Experts per token | 8 |
| Context length | 262,144 (config.json max_position_embeddings) |
| Languages | en, ko, es, de, ja, vi |
| Papers | arXiv:2601.01739 |
| License | k-exaone (other) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · ExaoneMoeForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · ExaoneMoeForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for K-EXAONE-236B-A23B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 569 GB · Multi-GPU node (<640 GB) |
|---|---|
| FP8 (1 byte/param) | 285 GB · Multi-GPU node (<640 GB) |
| INT4 (0.5 bytes/param) | 142 GB · Multi-GPU node (<640 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
All variants
16 variants
Lineage
quantized: 13finetune: 1adapter: 2
Top variants (16 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| LGAI-EXAONE/K-EXAONE-236B-A23B-FP8 | quantized | -FP8 | 237B | XXL | other | LGAI-EXAONE | 16 | 28.3k |
| furiosa-ai/K-EXAONE-236B-A23B-NVFP4A16 | quantized | — | 138B | XXL | other | furiosa-ai | 4 | 18k |
| LGAI-EXAONE/K-EXAONE-236B-A23B-GGUF | quantized | GGUF | — | — | other | LGAI-EXAONE | 19 | 11.2k |
| mradermacher/K-EXAONE-236B-A23B-i1-GGUF | quantized | GGUF | — | — | other | mradermacher | 0 | 8.9k |
| mradermacher/K-EXAONE-236B-A23B-GGUF | quantized | GGUF | — | — | other | mradermacher | 2 | 5.8k |
| Baekpica/K-EXAONE-236B-A23B-Mixed-Quant-GGUF | quantized | -Quant | — | — | other | Baekpica | 1 | 1.7k |
| Hyun9junn/K-EXAONE-236B-A23B-W4A16-G128 | quantized | -W4A16 | 33.1B | L | llama3 | Hyun9junn | 1 | 1.4k |
| Baekpica/K-EXAONE-152B-A23B-REAP-80E-BF16 | finetune | -BF16 | 151B | XXL | other | Baekpica | 1 | 1.3k |
| mkreynolds/K-EXAONE-236B-A23B-GGUF-4bit | quantized | GGUF | — | — | other | mkreynolds | 1 | 553 |
| haanjack/K-EXAONE-236B-A23B-MXFP4 | quantized | MXFP4 | 130B | XXL | other | haanjack | 1 | 526 |
| mlx-community/K-EXAONE-236B-A23B-8bit | quantized | -8bit | 237B | XXL | other | mlx-community | 0 | 493 |
| DevQuasar/LGAI-EXAONE.K-EXAONE-236B-A23B-GGUF | quantized | GGUF | — | — | — | DevQuasar | 0 | 408 |
| mlx-community/K-EXAONE-236B-A23B-6bit | quantized | — | 237B | XXL | other | mlx-community | 0 | 379 |
| mlx-community/K-EXAONE-236B-A23B-5bit | quantized | — | 237B | XXL | other | mlx-community | 0 | 347 |
| tazza26/K-EXAONE-236B-SFT-lora-r4 | adapter | — | — | — | — | tazza26 | 0 | 30 |
| gsjang/K-EXAONE-236B-SFT-lora-r4 | adapter | — | — | — | — | gsjang | 0 | 24 |
Repository changelog
18 commits
Repository changelog 18 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-08-03 | Update config.json | 61e6d57 |
| 2026-05-15 | Update README.md | f09edc1 |
| 2026-05-15 | Update news for FuriosaAI's NVFP4+GPTQ weights | d0ecfa0 |
| 2026-02-24 | Add redundant is_moe_layer for inference engines | 90d2d50 |
| 2026-02-06 | Update README.md and config.json for Transformers v5 | f7db87a |
| 2026-01-26 | Update README.md | adc21ca |
| 2026-01-22 | Update config.json | 01ae2bf |
| 2026-01-14 | Update config.json | 01a1c19 |
| 2026-01-13 | Update blog | e027c99 |
| 2026-01-12 | Update README.md | 6c389f0 |
| 2026-01-09 | Update README.md and config files | 27c3a76 |
| 2026-01-06 | Update technical report Arxiv link | 5aa6275 |
| 2026-01-04 | Update README.md | 1d92ab7 |
| 2026-01-03 | Update README.md | 2159cb4 |
| 2026-01-03 | Update README.md | b004998 |
| 2026-01-02 | Update technical report, API, and evaluation results | 3873da9 |
| 2026-01-02 | Fix eos_token_id | b0bf986 |
| 2025-12-31 | Initial commit | 636bca1 |
Papers
1 paper
Papers 1
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
- K-EXAONE Technical Report this model
Similar models
3 models
Similar models
- Qwen3-235B-A22B · 235B same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
- MiniMax-M2.5 · 229B same size band (XXL), also MoE, same task (text-generation), more providers
- Qwen3-235B-A22B-Instruct-2507 · 235B same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:14 UTC · Last verified 2026-10-25
Follow K-EXAONE-236B-A23B
Watch this model
- /feeds/models/LGAI-EXAONE--K-EXAONE-236B-A23B.xml(RSS 2.0) · every change to K-EXAONE-236B-A23B
- /feeds/families/LGAI-EXAONE--K-EXAONE.xml(RSS 2.0) · every change across the K-EXAONE family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.