Solheim AI
Solheim AI is a paid, EU-hosted service that rents private large language model (LLM) instances rather than metering tokens. Each project maps to one "virtual private LLM" (VPL): a dedicated endpoint and API key whose capacity is set by two dials — the number of concurrent in-flight requests (instances) and the context window each request receives. Inside that purchased size, customers can run as much as they like, with no rolling usage window, no reset timer, and no per-token overage billing; requests beyond the instance count queue instead of returning 429 errors. Pricing is a flat monthly fee, with the published entry point at EUR 15/month for one instance and a 64k context window, before VAT. The service exposes an OpenAI-compatible API, so existing tooling and SDKs — editors, coding agents, and custom backends using a base URL plus key — can point at the endpoint without code changes. Available open-weight models include Qwen3.6-35B-A3B (mixture-of-experts, 256K context, Apache 2.0), Qwen3.8-27B (dense, 256K context, Apache 2.0) and DeepSeek V4 Flash (256K context, MIT), with models pinned per instance and swappable by changing one string. Customers can also request additional open-weight models to be added to the library. Sovereignty is presented as an architectural property: GPUs run at EU-owned providers in EU datacenters with no US hyperscaler in the serving path, prompts are not used as training data, and requests never leave the region. Larger configurations, including dedicated GPUs or MIG slices with 256K windows, private networking, signed DPAs and invoiced billing, are offered as an enterprise tier. A separate unlimited coding plan is marketed for coding-agent users.
- Category
- LLM
- Pricing
- Paid
Problem solved
Token-metered LLM APIs create unpredictable costs and hard limits: usage windows reset, requests fail mid-task, and capacity is throttled exactly when a project needs it. Solheim AI replaces per-token billing and rate-limit resets with a flat monthly fee and a simple capacity model of concurrent instances plus context window, so peak workloads never cause lockouts or surprise invoices while keeping data in EU jurisdiction.
Use cases
- Running coding agents and IDE assistants (e.g. VS Code BYOK) against a private, OpenAI-compatible endpoint
- Serving multi-agent or parallel test loops that need several concurrent requests without hitting resets or 429s
- Processing long documents and codebases with up to 256K context windows
- Building GDPR- and AI Act-sensitive applications that must keep prompts and data inside the EU
- Running open-weight models such as Qwen or DeepSeek with pinned versions for reproducible workloads
- Backing internal tools and prototypes where predictable flat monthly costs matter more than per-token elasticity
Audience
- Developers
- ML Engineers
- Startups
- Enterprise
Pros
- Flat monthly pricing with no token metering, usage windows, resets or overage charges
- Capacity is explicit and controllable: instance count plus context window per project
- OpenAI-compatible API means no code rewrites for existing SDKs and agents
- EU-only hosting, permissive open-weight models, and prompts not used for training
- Requests beyond the instance count queue rather than failing with 429 errors
Cons
- Only a limited set of models is available, and additional models require requesting them
- Entry-level shared capacity is capped at 128K context; 256K requires dedicated enterprise pricing
- No free tier or trial mentioned, and pricing excludes VAT
- No GitHub repository or self-hosting option — it is a managed service only
- Detailed latency, throughput and SLA figures for shared instances are not published
Official website: https://solheim.ai/
Canonical LambdaBase page: https://www.lambdabase.com/tools/solheim-ai