Minerva
Validated 0.2, from-scratch 1.0, and Apollo — a Korean & code-specialized LLM evolving on three tracks
Azwell AI develops Korean & code-specialized LLMs on three parallel tracks. Minerva 0.2, built on an open-source 80B MoE base, has completed validation of our full training and evaluation pipeline (measured benchmark results published); on top of those assets we are developing Minerva 1.0, a from-scratch foundation model; and in parallel we are training Apollo, a flagship that combines NVIDIA Nemotron Super 120B with Korean-language training and coding RL. Beyond fine-tuning to full foundation models — internalizing the entire stack with our own technology.
Model Lineup — Three Tracks in Parallel
Minerva 0.2 — Validated
Open-source 80B MoE base + Korean CPT/SFT + our in-house Harness inference — training/evaluation pipeline fully validated with measured benchmark results published
Minerva 1.0 — In Development
A from-scratch pretrained foundation model — tokenizer, architecture, and data pipeline all designed in-house from a blank slate
Apollo — In Training
NVIDIA Nemotron Super 120B base + Korean-specialized training + coding RL post-training — a flagship aimed at top-tier performance
The three tracks are a staged technology-internalization strategy. With 0.2 we battle-tested data curation, training infrastructure, and our evaluation harness; on that validated pipeline we build the fully-controlled from-scratch 1.0; and in parallel, Apollo secures immediate top-tier performance by combining reinforcement learning with the strongest open base available (Nemotron Super 120B). Risk is divided across tracks — technology assets are shared between them.
Minerva 0.2 — A Validated Starting Point
Training + Harness Combined
Korean-language training combined with Harness inference techniques, with gains verified through real measurements — +6 points confirmed on Korean general knowledge
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
Minerva 0.2 Benchmark Results — Measured
| Benchmark | Minerva v0.2 |
|---|---|
| KMMLU (Korean knowledge) | 66 |
| HAE-RAE (Korean culture & common sense) | 64 |
| HumanEval (code) | 85 |
| MBPP (basic code) | 78 |
| HumanEval-X-ko ★ (Korean code) | 70 |
| KLUE-NLI (Korean reasoning) | 89 |
Domestic positioning — Rather than chasing the 'No.1 on general-knowledge benchmarks' race, Minerva 0.2 is a specialized model focused on real-world accuracy and realistic operating cost for handling code in Korean. It is strong on code (HumanEval 85) and Korean reasoning (KLUE-NLI 89), and — as an 80B-class model that runs on a single GPU (INT4) under a commercially unrestricted Apache 2.0 license — it offers the most practical alternative for Korean companies building in-house Korean coding assistants and RAG without large infrastructure. The general-knowledge dimension is reinforced through the from-scratch 1.0 and flagship Apollo tracks.
The Harness Methodology — Our In-House 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. Minerva 0.2's real-world validation confirmed a +6-point gain on Korean general knowledge — the first measured case in Korea to combine Korean-language training with the Harness approach. The same evaluation harness is reused for quality validation of 1.0 and Apollo.
Minerva 0.2 — 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
Minerva 1.0 — From-Scratch Foundation Model (In Development)
Designed From a Blank Slate
Tokenizer, model architecture, and training recipe designed in-house from scratch — internalizing foundation-model construction beyond open-source fine-tuning
Korean & Code-Centric Data Pipeline
The entire pretraining corpus pipeline — collection, cleaning, deduplication, quality filtering — built in-house
Inheriting 0.2's Validated Assets
The training infrastructure and evaluation harness battle-tested on 0.2 apply directly — structurally reducing from-scratch risk
Full-Stack Control
Complete control over data provenance, licensing, and model behavior — the foundation for customization in regulated industries like finance and the public sector
Apollo — Flagship on Nemotron Super 120B (In Training)
A Top-Tier Open Base
NVIDIA Nemotron Super 120B — starting from among the strongest reasoning performance of any open base available today
Korean-Specialized Training
CPT and SFT on Korean knowledge, culture, and domain data — directly closing the Korean-language gap of global base models
Coding RL Post-Training
Reinforcement learning from verifiable rewards (RLVR) such as test passes and execution results — optimized for code that actually runs, not just human preference
Why 'Apollo'
Following Minerva, goddess of wisdom — Apollo, god of reason, logic, and foresight: the name of a flagship aimed at both Korean and code
Roadmap — From Validation to Foundation
Done · Minerva 0.2
Open-source 80B MoE + Korean CPT/SFT/DPO + Harness — measured benchmark results published
In Training · Apollo
Nemotron Super 120B + Korean training + coding RL — securing flagship performance
In Development · Minerva 1.0
From-scratch pretraining — a fully in-house foundation model
Expansion · Multimodal
Vision-language expansion under review once a VLM base is secured
Training pipeline: CPT (Korean continued pretraining) → SFT (domain supervised fine-tuning) → DPO (preference alignment) → RL (reinforcement learning from verifiable rewards, Apollo). Minerva 0.2's measured benchmark results are published, and model weights, code, and evaluation results are released as open source on HuggingFace and GitHub (license: Apache 2.0).