CUECLOUD / OPEN MODEL EXPLORER
Exceptional models. Available to run on your infrastructure. Explore the capability, compare the evidence, and find the right starting point for Vault.
Each axis shows its original 0–100 score. Shapes show a profile, not a combined ranking. Click a benchmark below the chart to inspect it.
01 / THE EVIDENCE
Switch the workload and benchmark. See where open models lead, where they’re close, and where closed models still have an edge.
DEEPSEEK / PUBLISHED RESULTS
V4.1 Flash leads this published Terminal-Bench 2.1 comparison. Switch benchmarks to explore its strengths and tradeoffs.
Completing terminal tasks.
SCORE / HIGHER IS BETTERDeepSeek Harness Minimal, 1M context, reasoning effort 100. Publisher-reported frontier comparison.
Publisher-reported results, not CueCloud measurements. Quantization, hardware, context, and agent harness can change performance. This is a selected comparison, not a universal ranking.
Source reviewed September 23, 2026. Full methodology ↗
EXPLORE / CAPABILITY PER MODEL SIZE
Click a point to inspect a model. Compare total parameter count with published benchmark scores from Qwen’s release table.
Publisher-reported snapshot. Parameter count is a size measure, not a prediction of speed, memory use, or deployment support. The closed model’s parameter count is not disclosed here, so it appears as a score reference rather than a point.
02 / THE MODEL CATALOG
A curated selection for private deployment. We confirm model, runtime, license, and hardware compatibility with you before installation.
Image and text understanding with 8B active parameters during prefill and 16B during decode. The checkpoint also includes conditional memory and vision components.
Plan for substantial memory and a larger deployment. Active parameters describe compute per token, not the amount of weights you need to store.
“Open” here means downloadable model weights. License terms vary by model. Catalog inclusion is a deployment option to assess, not certification on every Vault configuration.
03 / YOUR HARDWARE
Move the memory budget and change precision. Get a feel for model scale before we size your actual deployment.
Raw weights = total parameters × bits ÷ 8. Decimal GB. Runtime, KV cache, quantization metadata, and operating-system memory are extra. A weight budget is not a serving guarantee; precision support varies.
How Vault is configured ↗04 / WHY BRING IT INSIDE?
Run selected models next to your code and research. In a validated air-gapped deployment, inference does not require a public model API.
Select a checkpoint, test it on your work, and keep that version in service. Add a different model when your needs change.
Connect your repositories and tools through Blitz. We configure the model, inference stack, and engineering workflow together.
Tell us what your engineers and researchers need to do.
We’ll put the right models on the right hardware.
