Made in Korea is not best at Korean: five budget LLMs on four Korean tests

한국 모델이 한국어를 제일 잘하진 않았다: 저가형 LLM 5종, 한국어 테스트 4종

On Korean linguistics, NER, legal text and instruction following, DeepSeek V4.1 Flash and GPT-6 Luna lead. Upstage's Solar models trail, and the newer Solar Mini 4 is last on all four tests.

#llm#korean#benchmark

Contents

I ran five budget-tier LLMs on four Korean tests. The short version: the two Korean-made models are not the best at Korean. DeepSeek V4.1 Flash and GPT-6 Luna lead. Upstage’s Solar Pro 4 is significantly behind DeepSeek on all four tests, and Solar Mini 4 is last on all four.

The numbers are kept current on the benchmark page. This post explains what was run and how to read it.

TL;DR

ModelKoBALTKLUE-NER (F1)LBoxHanIFEvalCost, all four
DeepSeek V4.1 Flash81.7%78.684.6%97.2%$8.23
GPT-6 Luna82.4%74.884.5%93.2%$1.00
GLM 5.3 Flash71.0%76.480.7%95.3%$10.22
Solar Pro 472.7%76.481.4%93.0%$3.61
Solar Mini 459.1%65.177.0%92.5%$3.66

Reasoning on, one run per model, measured 2026-09-29 to 10-03. Bold marks the gaps this post is about, not a winner.

  • DeepSeek V4.1 Flash is never significantly beaten.
    • It ties GPT-6 Luna on KoBALT (p = 0.66) and LBox (p = 1.0).
    • On NER it is ahead of every other model: by 2.2–3.8 F1 over GPT-6 Luna, GLM and Solar Pro 4 (p ≤ 0.0042), and by 13.5 over Solar Mini 4.
    • On HanIFEval it is ahead of GPT-6 Luna and Solar Pro 4 (p ≤ 0.0015).
  • GPT-6 Luna matches DeepSeek on knowledge for an eighth of the cost: $1.00 against $8.23 for all four tests. It loses on format: it adds a final period when asked to copy a sentence, and it ignores a requested *** paragraph divider.
  • The Korean models trail.
    • Solar Pro 4 is significantly behind DeepSeek on all four tests, and behind GPT-6 Luna on KoBALT and LBox.
    • Solar Mini 4, listed on OpenRouter on 2026-09-23, is last on all four. It is significantly behind every other model on KoBALT, NER and LBox.

What was compared

All five models sit in the same OpenRouter price band, checked on 2026-09-29 (Solar Mini 4 on 2026-10-03).

ModelFromServed byPrice in / out (per 1M)
DeepSeek V4.1 FlashDeepSeek (CN)StreamLake, fp8$0.15 / $0.60
GPT-6 LunaOpenAI (US)OpenAI$0.10 / $0.50
GLM 5.3 FlashZ.AI (CN)Z.AI, fp8$0.15 / $0.50
Solar Pro 4Upstage (KR)Upstage$0.09 / $0.36
Solar Mini 4Upstage (KR)Upstage$0.05 / $0.20

The setup:

  • Each model is pinned to one provider with fallbacks off.
  • Reasoning is on at the provider’s default effort, with temperature 0 and a 64,000-token output limit.
  • Each item is answered once.

The DeepSeek numbers are for a third-party fp8 deployment. DeepSeek’s own endpoint was not available to my account.

I also dropped two Gemma 4 models:

  • Gemma 4 31B’s only working full-precision endpoint was too rate-limited to finish a run.
  • Gemma 4 26B kept reasoning until it hit the token limit on most KoBALT and NER items.

Four tests, four different skills

  • KoBALT (SNU, 2025): 700 expert-written ten-option questions on Korean linguistics. The questions cover syntax, semantics, pragmatics, phonology and morphology. This is knowledge about the language, not fluency in it.
  • KLUE-NER: 1,000 dev sentences from news and movie reviews. The model copies each sentence and marks its people, places, organizations, dates, times and quantities. A copy that changes the sentence scores as “no entities found”.
  • LBox Open casename: given the facts of a Korean court judgment, pick its case name from 100 labels.
  • HanIFEval: 429 verifiable instructions in Korean, such as word counts, forbidden words, formats and “repeat the request first”. Code scores them; no judge model is involved.

I built HanIFEval for this study because existing Korean IFEval translations break their own checkers. In one audited set, 27 of 342 Korean prompts said something different from what the checker scored. For example, “more than 2 times” was translated as “2번 이상” (2 or more) while the checker counted 3 or more. HanIFEval translates each prompt together with the checker’s arguments, so the two cannot drift apart. After two external reviews it was revised to v1.1; the scores here are on v1.1. How it was built is in a separate post.

Results

KoBALT and LBox split the original four models into the same two tiers:

  • DeepSeek and GPT-6 Luna are on top.
  • GLM and Solar Pro 4 are below them.

All 8 tests between the tiers are significant, and none of the 4 within a tier is.

Solar Mini 4 falls below both tiers: 59.1% on KoBALT is 13.6 points under Solar Pro 4.

GLM’s KoBALT score is about thinking time, not knowledge. 91 of its 700 replies were still reasoning at 64,000 tokens. Those count as wrong. On the 609 questions it did answer, GLM is 81.6% correct, level with the top tier.

On LBox, the case name appears word for word in the facts in 282 of the 1,000 items, and every model scores 89–94% on those. The models separate on the other 718.

NER: DeepSeek leads, and format matters

DeepSeek’s F1 is above every other model’s. GPT-6 Luna’s lower score is mostly formatting: in 73 of its 79 failed replies, the copy differed from the sentence only by an added final period or similar trailing punctuation. Forgive that, and GPT-6 Luna cannot be separated from GLM or Solar Pro 4.

Solar Mini 4 failed to copy 159 of 1,000 sentences. It dropped words, repeated words and changed particles.

Instruction following: near ceiling

All five models score 92.5–97.2% on HanIFEval, and 351 of the 429 items are passed by all five. The only significant gaps are DeepSeek over GPT-6 Luna, Solar Pro 4 and Solar Mini 4.

The per-type numbers say more than the total:

  • Solar Pro 4 wrote too little on 13 of 43 length instructions.
  • GPT-6 Luna missed the requested *** paragraph divider on 9 of 22 items.
  • GLM failed 7 of 45 forbidden-word instructions, often by restating the words while saying it had avoided them ("‘변경’과 ‘여유’라는 단어는 사용하지 않았습니다").

Caveats

  • One configuration per model: one provider, default reasoning effort, one sample per item. Run-to-run variance is not measured.
  • Contamination is not controlled.
    • KLUE (2021), LBox (2022) and KoBALT (2025) were public with answers before these models were released.
    • HanIFEval is new, but its English source, IFEval, is from 2023.
  • Non-answers count as wrong. That costs GLM most on KoBALT, and costs GPT-6 Luna on NER.
  • Five models is a handful. “The Korean-made models trail” describes these two Solar models, not Korean LLMs in general.
  • HanIFEval is model-validated, not human-validated. Its translator was Gemini; its reviewers were Claude and GPT-6.
  • Statistics. McNemar exact tests (a paired bootstrap for NER) with Bonferroni over the 6 pairs of the first four models. Solar Mini 4 joined later and is tested in its own family of 4 pairs.
  • Cost covers the scored runs only, at OpenRouter prices on the run dates.

Reproducing

Everything is in developer0hye/llm-evals:

  • the harness and the pinned dataset revisions;
  • per-item logs for every model, and the analysis scripts;
  • a research log for each study, covering every decision, dead end and p-value behind the claims above.

The HanIFEval dataset, its scorer and its changelog are on Hugging Face (tag v1.1).