Migrate to the v0.11.0 lineup

With v0.11.0 the Elevarq product lineup is simplified to four plans, four runtime variants — and two CPU variants are retired: medium-cpu (the 8B model on CPU) and large-cpu (the 30B model on CPU). This guide covers why, who is affected, and exactly how to move.

The lineup at a glance

Only the Starter model runs on CPU; every larger model is GPU-only. Qwen3-30B-A3B is a Mixture-of-Experts model (~3B parameters active per token) — 30B-class quality at high throughput, not a dense 30B.

Why the change

Running the mid-size models on CPU added cost without product benefit. In validation, the 8B model on CPU was slower than the 4B on the same hardware — over 45 minutes for some reports — with no reliable quality gain, while the GPU variants deliver the same or stronger analysis in seconds to tens of seconds (14B: ~30 s per snapshot; 30B-A3B: ~40 s, ~102 tokens/s, both validated on an A10G). Retiring the CPU mid-tiers removes the worst-value configurations from the lineup.

Who is affected

• You run medium-cpu (8B on CPU) → your plan (Professional) now runs medium-gpu (14B). You move to a GPU host.
• You run large-cpu (30B on CPU) → your plan (Business) now runs large-gpu. You move to a GPU host.
• You run small-cpu, medium-gpu, large-gpu, or xlarge-gpu → nothing changes; no action needed.

How to migrate

The move is a variant swap on a GPU host — the same in-place procedure as any plan upgrade, and your data is preserved throughout:

• Provision a GPU host for your variant — sizing, VRAM minimums, and validated instance types are in the sizing guide. (medium-gpu: ≥16 GiB VRAM, e.g. g5.xlarge; large-gpu: ≥24 GiB VRAM and 32 GiB host RAM, e.g. g5.2xlarge.)
• Back up first, as for any upgrade — see the release-upgrade guide.
• Swap the image variant (medium-cpu → medium-gpu, large-cpu → large-gpu) keeping the same persistent volumes and the same master key. Registered databases, analysis history, reports, findings, users, and settings all carry across unchanged.
• Verify: the Insights page shows the GPU variant being served; reports enrich in seconds rather than minutes.

The move is never automatic. Elevarq never silently migrates a CPU deployment to GPU — you provision the host and perform the swap. Until you do, a deployment still serving a retired variant keeps working but shows a migration warning in Workbench naming the replacement variant.

Licenses

• Existing licenses for current variants are unaffected.
• A license that requests a retired variant (medium-cpu / large-cpu) is refused at runtime with a clear "no longer supported" error naming the replacement — never an ambiguous model-load failure, and never a silent substitution. Contact us and we reissue the license for the GPU variant of your plan at no plan change.

Database limits and add-ons

The included database counts are 1 / 3 / 25 / contract-defined (see the table above). Two add-ons extend any plan:

• Additional databases — raise your plan's database count without a plan change.
• Parallel GPU concurrency (Professional and Business) — one flat unlock that lets a deployment analyze several databases at the same time on one GPU. It is bounded by the GPU's memory and never metered per database.

Current add-on pricing is on the pricing page.

Entitlement vs capacity

Your plan's database count is a commercial entitlement — how many databases your license admits. It is not a claim about raw technical capacity, and we deliberately make no "thousands of databases on one node" promises: the numbers we publish are the configurations we validate. If your estate is larger than the Business plan's 25 databases, that is the Enterprise conversation — scale is sized and contracted explicitly.