CUSTOM MODEL
Keural Cortex 8B v1.0
Korean–English · instruction and tool tuning
REPRODUCIBLE MODEL EVALUATION
Cortex v1.0 Comparative evaluation
Seven-category short evaluation · NVIDIA DGX Spark · BF16
Available when all jobs complete
Connecting to Spark
— / —jobs completed
—
Fetching live evaluation state…
Operator controls · private access
GPU utilization—NVIDIA GB10
Unified system memory—Capacity is not a speed indicator
Elapsed wall time—Live updates every 5 seconds
ⓘ Small exploratory samples. No overall winner established. Differences are percentage points, not significance tests.
Audited development validation · Base-only comparison
Same requests, schemas and output budget; explicitly different serialization.
COMPARATOR
Qwen3-8B-Base
Original pretrained baseline
Side-by-side results
Pending results stay blank. No results are inferred or filled from published leaderboards.
| Benchmark / subtask | Cortex v1.0 | Qwen3-8B-Base | Difference | Coverage / metric |
|---|
Standard tasks use matching completion prompts. Chat, tools and context use the same plain transcript on both models. Actual input lengths reserve output space. 64K is a Cortex-only extension, not a paired win. Synthetic probes are not standardized capability certifications.
Coverage across seven categories
Frozen standard samples and bilingual enterprise fixtures; 64K is a Cortex-only extension. Download public audit pack
Live activity
Current or most recent evaluation job
Loading diagnostic output…
Methodology & reproducibility
Keep the comparison protocol visible alongside the scores.
Matched completion
Standard benchmarks use identical raw completion prompts, few-shot counts and frozen sample indices on both models. Up to 50 samples per leaf task; budgeted coverage, not complete test splits.
Common-format behavior
Cortex uses its native template. Qwen3-8B-Base uses a plain transcript. Same requests and schemas, greedy decoding, 128-token cap and repetition penalty 1.0. No Instruct weights.
Interpretation limits
1,280 synthetic domain fixtures per model, plus context cases; not independently authored holdout. Schema, tool choice and limited argument checks are saved separately. No overall winner or statistical claim.
All jobs and evaluation settings
| Model / track | Job | State | Settings |
|---|