Keural Nova is a Korean-focused instruction-tuned model built by MKD on top of Qwen/Qwen3.6-35B-A3B (Apache 2.0), a 35B-parameter Mixture-of-Experts model with ~3B active parameters per token. Full credit for the base architecture and pretraining goes to the Qwen team at Alibaba Cloud — this release is a supervised fine-tune (LoRA, merged into the base weights) targeting Korean conversational quality, with an English/general replay mix to limit capability regression.

Version note: v1.0 is the end-of-epoch-1 checkpoint (7,500 optimizer steps), selected as the best held-out eval-loss point of the training run. A sibling release, Keural Nova v1.1, is the full 2-epoch checkpoint (15,846 steps) — see the comparison table below; the two make slightly different trade-offs and we publish both transparently.

Evaluation

All numbers measured by us with lm-evaluation-harness 0.4.12 (vLLM backend, bf16, seed 42, default few-shot settings), same harness and settings for base and fine-tuned models.

Benchmark	Suite	Base model	v1.0 (this model)	v1.1	Δ v1.0 vs base
KoBEST	Korean	62.88	68.82	67.29	+5.94
HAE-RAE	Korean	75.53	73.42	74.52	−2.11
KMMLU	Korean	64.46	62.52	62.38	−1.94
MMLU	General	83.83	82.07	82.37	−1.76
GSM8K (strict-match)	Math	32.15*	87.11	86.88	+54.96*
HumanEval (pass@1)	Code	62.20	20.73	11.59	−41.47

* The base model's GSM8K score is heavily depressed by answer-format mismatch under strict-match extraction; the post-SFT jump is largely format compliance (flexible-extract scores the fine-tune at 87.49). Treat the delta as a formatting improvement, not a ~55-point reasoning gain.

Honest summary: the fine-tune substantially improves Korean conversational/comprehension performance (KoBEST) and answer formatting, at the cost of small regressions on knowledge-recall benchmarks (KMMLU, HAE-RAE, MMLU) and a large regression on code generation — see Limitations. If your workload is Korean dialogue/comprehension, this model is the point of this release; if it is knowledge QA or coding, use the base model.

Intended use & limitations

Intended: Korean conversational AI, Korean text comprehension and generation, bilingual (ko/en) assistant use, RAG-style grounded answering in Korean.
NOT suitable for code generation. HumanEval drops from 62.2 → 20.7 pass@1. Root cause (identified, not speculative): a large fraction of code examples in one upstream training source had whitespace/indentation corrupted during that dataset's original export, and the model learned the corrupted formatting. A future release will address this with repaired data.
Knowledge-heavy benchmarks (KMMLU/HAE-RAE/MMLU) regress by ~1–2 points vs base.
Context length: SFT sequences were up to 8,192 tokens; the base architecture supports 262,144 tokens natively and up to 1,000,000 tokens via YaRN rope scaling. Extended-context serving up to 1M tokens has been run in production through MKD's OpenAI-compatible gateway (YaRN factor 4.0) and works. Note: long-context answer accuracy has not been measured with a dedicated benchmark (e.g., needle-in-a-haystack / RULER), so validate for your specific long-document workload.
The model inherits the base model's multimodal architecture (vision tower is present and untouched), but this fine-tune neither trained nor evaluated vision — treat it as a text model.

Training details


Base model	Qwen/Qwen3.6-35B-A3B @ `995ad96e`
Method	LoRA (rank 32, α 64, dropout 0.0), merged into base weights
Adapted modules	attention (`q/k/v/o_proj`), Gated-DeltaNet linear attention (`in_proj_qkv/z/b/a`, `out_proj`), shared experts, and all 256 routed experts per layer (fused `gate_up_proj`/`down_proj` tensors via PEFT `target_parameters`)
Deliberately frozen	MoE router (`mlp.gate`), `shared_expert_gate`, embeddings, LM head, vision tower
Data	253,528 examples (~85% Korean / 15% English replay), 13 sources, ChatML, empty-`<think>` (non-thinking) format
Schedule	effective batch 32, LR 1e-4 cosine, warmup 5%, bf16, DeepSpeed ZeRO-2, 2× H200
This checkpoint	step 7,500 (~1 epoch) — best held-out eval loss (0.857) of the run
Loss masking	assistant tokens only; empty think-block scaffold excluded from loss

Router-integrity and expert-targeting were verified programmatically before and after training (the router and gates are bit-identical to the base model in this release; the routed-expert weights are where most of the adaptation lives).

Training data & licensing

Sources (per-row license tags preserved from upstream): Korean conversation/RAG/QA and instruction data from the mkd-chanwoo keural datasets (CC BY 4.0 / MIT / Apache-2.0 / CC BY-SA 4.0 subsets only), plus eight sources from Mkd-Yonas/keural-SFT-chatml-ko-v1 (MIT / Apache-2.0 / CC BY 4.0), plus English replay data (ultrachat, openorca, alpaca, magicoder, mathinstruct, gsm8k — MIT/CC BY 4.0).

License mix of the final training set: CC BY 4.0 (105k rows), MIT (87k), Apache-2.0 (46k), CC BY-SA 4.0 (3.5k), unlabeled (13.7k — two Korean instruction subsets whose upstream rows carry license: unknown; upstream documentation states Apache-2.0, verification pending).

Excluded by policy: all AI Hub-derived subsets (non-commercial research terms) were filtered out of training data.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mkd-hossain/Keural-Nova-v1.0"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "한국의 전통 발효 음식에 대해 설명해 주세요."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True,
                                 enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.95)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

vLLM serving (this is a fully merged model — no adapter loading needed):

vllm serve mkd-hossain/Keural-Nova-v1.0 --tensor-parallel-size 2 --dtype bfloat16

The model is trained in the base model's non-thinking chat format (enable_thinking=False); thinking-mode behavior is inherited from the base model untouched.

Acknowledgements

Built on Qwen3.6-35B-A3B by the Qwen team (Alibaba Cloud), released under Apache 2.0. Korean datasets curated by the MKD data team. Fine-tuning, evaluation, and release engineering by MKD.