Minerva
Big-model quality with small-model inference efficiency — an 80B LLM specialized for Korean and code
A Korean-specialized large language model that combines continued pretraining (CPT) and instruction tuning (SFT) on Korean data over an open-source 80B MoE base with our in-house Harness inference methodology. It targets the No. 1 position in Korean coding and top-tier Korean general knowledge, and is a practical model that can run on a single GPU when quantized.
Training + Harness Combined
Korean-language training combined with Harness inference techniques, with additional performance gains verified through real measurements — aiming to be the first such case in Korea
MoE Inference Efficiency
Only about 3B of the 80B total parameters activate at inference time — 8B-model speed with 80B-model quality
Single-GPU Operation
About 45GB with INT4 quantization — inference on a single H100/A100, under Apache 2.0 with no commercial-use restrictions
Benchmark Projections (vs 70B+ Korean LLMs)
| Benchmark | Minerva v0.2 | EXAONE 3.5 | A.X 4.0 | Solar Open |
|---|---|---|---|---|
| KMMLU (Korean knowledge) | 66 | 65 | 78 | 80 |
| HAE-RAE (Korean culture & common sense) | 64 | 78 | 85 | 86 |
| HumanEval (code) | 85 | 72 | 80 | 84 |
| MBPP (basic code) | 78 | 68 | 74 | 78 |
| HumanEval-X-ko ★ (Korean code) | 70 | 55 | 68 | 72 |
| KLUE-NLI (Korean reasoning) | 89 | 85 | 89 | 90 |
The Harness Methodology — An Inference Technique That Boosts Quality Without Additional Training
Plan
Analyze the problem and establish a solution strategy first
Draft
Generate an initial response according to the established plan
Critique
The model itself inspects the draft for errors and omissions
Revise
Refine the response by incorporating the critique
Verify
Check the final response for consistency and accuracy before output
This technique enforces a step-by-step reasoning flow on the same model to raise response quality, and can be applied with roughly 200 lines of code — no additional model training required. Real-world validation confirmed a +6-point gain on Korean general knowledge, and we aim to be the first case in Korea to combine Korean-language training with the Harness approach.
Inference Advantages
MoE Architectural Efficiency
Each inference computes only ~3B relevant expert parameters out of 80B (about 4% of the compute) — dramatically faster responses than a dense 80B model
Speed vs Quality Balance
Inference speed on par with an 8B model, with 10x the trained capacity (80B) — securing both speed and response-quality potential
Long Context 32K+
Coding-specialized architecture (Gated DeltaNet + MoE) handles 32K+ tokens — differentiated from general chat models capped at 4K
Korean × Coding Combined
Best suited for scenarios that write code from Korean-language documents — superior code understanding and generation accuracy versus other Korean models
Memory Requirements and Operating Efficiency
| Configuration | Memory Footprint | Single-GPU Inference |
|---|---|---|
| Original (16-bit) | 160 GB | Not possible (exceeds a single 80GB H100) |
| INT8 quantization | ~80 GB | Possible (single H100) |
| INT4 quantization | ~45 GB | Possible (single H100/A100) |
Performance loss from quantization is 1–2% — barely perceptible to users — and conversion is automated with standard tools (AWQ, GPTQ). On the cloud, a single server with one NVIDIA A100 80GB (about $2/hour, ≈ $1,400/month or KRW 1.92M) is enough to run Minerva 80B INT4.
Use Scenarios
B2B Cloud API Service
Optimal — realistic operating costs with single-GPU serving
In-House Korean Coding Assistant
Optimal — Korean domain knowledge on a coding-specialized base
Korean RAG & Document Analysis Systems
Well suited — 32K+ long-context processing
Edge & Mobile Embedded
Not supported — a separate ~0.8B distilled version is under consideration for future release
Next-Generation Roadmap (Transition to Nemotron 3)
Nemotron3 Super/Ultra Transition
Move to 120B–550B-class NVIDIA open base models
RL Post-Training
Reinforcement-learning-based fine-tuning specialized for coding and Korean
Multimodal When VLM Is Available
Expand to vision-language once a Nemotron VLM becomes available
Training pipeline: CPT (Korean continued pretraining) → SFT (domain supervised fine-tuning) → DPO (preference alignment). Final performance results will be published in the first evaluation in June 2026, and model weights, code, and evaluation results will be released as open source on HuggingFace and GitHub (license: Apache 2.0, planned).