SciCode
DeepSeek V4 FlashScore change+14.44 percentage points
v1
QUANTUMMIND
Conceptual visual
OUR FIRST MODEL
QuantumMind v1
Built under QART.
Exploring new possibilities in reasoning.
QuantumMind v1 is our first model built under the QART methodology. Public technical details and evaluation results will be presented in the technical report.
01 / METHODOLOGY
QART and QuantumMind v1
QART is our framework and methodology. QuantumMind v1 is the first model built under it. Technical descriptions follow the public report.
QART / CONCEPTUAL OVERVIEWQUANTUM-AUGMENTED REASONING
CONTEXT
Task and context
Understand context and constraints
REASONING
Quantum-enhanced reasoning
Explore methods for complex reasoning
EVALUATION
Results and evaluation
Evaluate methods through experiments
02 / EVALUATION
Results and comparisons
Model comparisons and method experiments are presented separately, each with its own evaluation context.
QuantumMind v1 · Model comparison
Comparable evaluation results and methodology will accompany the technical report.
QART enhancement across models
Internal paired evaluations · Method experiments, not a QuantumMind v1 model comparison
| BENCHMARK | DeepSeek V4 Flash | GLM-5.3 |
|---|---|---|
| τ²-Bench | +4.0% | +5.9% |
| τ³-Bench | +47.6% | +4.8% |
| SciCode | +84.1% | +10.1% |
| LHTB | +35.9% | Not provided |
| DeepSWE | −7.8% | +19.1% |
| Terminal-Bench 4.0 | +44.4% | +44.4% |
This page presents 11 internal paired results across 2 models and 6 benchmarks: 10 improved and 1 declined.
How the metrics work
Relative change = (score with QART − Direct score) / Direct score × 100%.
Direct is the baseline without QART. LHTB uses its native reward metric. A missing result does not mean a score of zero.
Results and limitations
DeepSeek V4 Flash declined by 7.8% on DeepSWE. Gains depend on the model and task.
Results may differ from official model scores and do not constitute a commitment to production performance.
Source: existing internal evaluation materials, to be checked against the technical report.
Terminal-Bench 4.0
GLM-5.3Score change+6.34 percentage points
03 / READING THE RESULTS
Understanding the results
Define the comparison
Current paired experiments compare the same model before and after adding QART. QuantumMind v1 model comparisons will be presented separately.
Distinguish the metrics
Relative improvement, actual scores, and percentage points are labeled separately. Metrics from different benchmarks are not combined into one overall score.
Clarify the scope
Negative results and missing values within this page’s scope are retained. Public experimental conditions and limitations follow the technical report.
TECHNICAL REPORT
The technical report.
Coming soon.
QuantumMind v1 · Public methodology and evaluation
Download available when published
RESEARCH & COLLABORATION