Pick your GPU — or enter your VRAM — and estimate which open-weight models could fit, at which quantization, with estimated headroom. These are self-hostable candidates. Sovereignty requires deployment evidence for identity, network access, telemetry, artifacts and recovery.
Reviewed cache configurations: 2026-10-29. Select an exact model or compare the curated set. Legacy candidates use a rough parameter-based cache estimate. Specialized architectures are listed separately below.
Held free after the OS/apps deduction. Working budget: 21.6 GB. This reserve is separate from the 4 GB runtime allowance.
Tokens include prompt, retrieval, tool history and generated output at peak. Count simultaneously retained sequences, not registered users or daily requests. Custom memory should already exclude memory occupied by other applications.
Verify the selected device backend, model architecture, weight format and quantization kernels in the pinned runtime.
Catalog additions reviewed 27 October 2026. These models are not included in the automatic fit count. Hybrid state, mixed attention, multimodal encoders and specialized quantization need artifact-specific measurements. Published context limits do not guarantee a fit on your selected hardware.
128K configured context · Model-specific sizing required
Apache-2.0. MoE with mixed sliding/full attention. MXFP4 expert weights are not interchangeable with a generic Q4 estimate; verify backend support, cache allocation and Harmony formatting.
Official model card · reviewed 2026-10-27128K configured context · Model-specific sizing required
Apache-2.0. Budget all resident experts, mixed-attention cache and runtime overhead. Publisher memory examples are not guarantees at maximum context or concurrency; benchmark the exact artifact.
Official model card · reviewed 2026-10-27256K native context · Model-specific sizing required
Apache-2.0. Hybrid DeltaNet/attention with a vision encoder. Extended context requires explicit configuration. Benchmark cache, recurrent state and image memory for the chosen artifact.
Official model card · reviewed 2026-10-27256K context · Model-specific sizing required
Apache-2.0. Dense multimodal model. Mixed attention and image inputs require cache and encoder budgeting beyond a parameter-only estimate.
Official model card · reviewed 2026-10-27256K context · Model-specific sizing required
Apache-2.0. Budget resident weights for the full MoE, not only active parameters. Validate attention cache and multimodal overhead on the selected runtime.
Official model card · reviewed 2026-10-2764K context · Model-specific sizing required
OpenMDW-1.1. 8B total / 4B effective parameters. The 4GB artifact label is not a total serving-memory guarantee: add cache, runtime and any speculative drafter.
Official model card · reviewed 2026-10-27Up to 256K context · Model-specific sizing required
OpenMDW-1.1. Quantized MoE. The 24GB label is not a fit guarantee at arbitrary context or concurrency. Verify IQ2 kernel support and the exact runtime/hardware requirements.
Official model card · reviewed 2026-10-271M published context · Model-specific sizing required
MIT. Asymmetric encoder-decoder architecture and compressed cache need dedicated sizing. Active parameter counts do not represent resident weight memory. Validate long-context capacity separately.
Official model card · reviewed 2026-10-27Highest weight precision that meets your budget is shown; this does not rank measured quality. Lower memory estimates are listed first. Runtime examples below are not verified for your selected hardware. Check CUDA, ROCm or Metal support, quantization kernels and per-device sharding before deployment.
Formula: weights × format allowance + cache + 4 GB runtime, within the budget after reserve. Weight estimates include a 15% packing buffer; cache uses 2-byte elements independently of weight quantization. Figures use decimal GB and approximate catalog capacities; confirm actual available bytes. Reviewed configurations calculate K and V from layers × KV heads × head dimension × tokens × sequences.
| Model | License | Context | Quantization | Memory breakdown | Headroom | Runtime examples |
|---|---|---|---|---|---|---|
Qwen3-8B Model card: reviewed 2026-10-20Native 32K context; extended context is not assumed. Quantized artifact and runtime support require validation. Cache configuration · reviewed 2026-10-29 | Apache-2.0 | 32K | INT8 | 14.6 GB Weights 9.4 GB | 39% | Compatible local runtime; verify release |
Download your results above without an email address. This form sends a subscription request to SovAIHub for occasional deployment notes; it does not automatically email your results.
Save your own GPU, model and workload settings. Review the architecture-aware capacity estimate, then download a versioned vLLM pilot package for a supported model.
Open deployment workspaceWork the estimate the other way: start from a model class and size the hardware you need.
Sovereignty tiers, deployment runtimes, and a decision guide for private AI model selection.
Get a fixed-scope architecture review: model selection, hardware sizing, runtime topology, and a written findings document for your private RAG or air-gapped AI platform.
Estimates are approximate. Actual memory depends on runtime, batching, quantization format, KV cache implementation, and model architecture. Always benchmark before purchasing hardware.