In development — early access opens with Phase 1

Feldspar

Start any open-source model with no server of your own. Plug in the guardrails, tools, and RAG connectors you need from the marketplace.

Unified APIshell
curl $FELDSPAR_URL/v1/chat/completions \
  -H "Authorization: Bearer $FELDSPAR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-3.3-70B-Instruct",
    "messages": [
      { "role": "user", "content": "Summarize this ticket." }
    ]
  }'

Same request, different model

  • open"model": "meta-llama/Llama-3.3-70B-Instruct"
  • open"model": "Qwen/Qwen2.5-72B-Instruct"
  • closed"model": "anthropic/claude-opus-5"

One parameter changes. Nothing else in your code does.

Request early access →Gneiss Labs
01 — Hosting

Three ways to run a model

You pick the trade-off between cold starts and idle cost. Feldspar does not rent you a dedicated GPU unless you ask for one.

ModeBehaviorYou pay
SharedPopular models stay loaded on shared GPUs. No cold start.Per token
Serverless dedicatedThe model loads on demand and scales to zero. Cold starts occur.Per GPU second
Always-on dedicatedThe GPU stays on. For steady volume or privacy requirements.GPU cost plus markup

LoRA adapters run against a shared base model, so many adapters sit on one GPU. Rates are not published yet — see the roadmap below.

02 — Request path

What happens to a request

Plugins run inside the request pipeline, not beside it. Every step is one you can see, time, and turn off.

  1. Gatewayauth · rate limit · balance
  2. Input guardrailsplugin
  3. RAGplugin
  4. Modelopen or closed
  5. Toolsplugin
  6. Output guardrailsplugin

Third-party plugins run in a sandbox and declare their permissions: network access, data access, and cost per call. The dashboard shows the latency and cost each one adds.

03 — Marketplace

Harnesses, not glue code

The parts everyone rebuilds — a PII filter, a retrieval connector, an eval — attach to a deployment instead of living in your application.

  • Guardrails

    PII filter · prompt-injection filter · content safety · topic limits

  • Tools

    Web search · database query · code run · MCP servers

  • RAG connectors

    Google Drive · Notion · S3 · websites · SQL

  • Evaluators

    Hallucination check · groundedness score · regression tests

  • Harness templates

    Prompt, tools, guardrails, and RAG preassembled into one setup

First-party plugins ship before the marketplace opens to other publishers. Tools use MCP rather than a proprietary interface.

04 — Pricing

Three tiers, metered per unit

Metering is per token for inference, per second for GPU time, per call for plugins, and per GB for RAG storage and embeddings.

  • BYOK

    Connect your own provider key and pay the provider directly. Feldspar charges a flat monthly fee for the software layer.

  • Managed

    Buy prepaid credits. Feldspar supplies model access and meters what you use.

  • Enterprise

    Invoiced billing, custom contract, private deployment.

Rates are not set. We would rather publish nothing than publish a number we intend to change.

05 — Status

What exists, and what does not

Feldspar is being built in the open. This is the whole roadmap, not the parts that sound finished.

  1. Phase 1Building now

    5 to 10 open-source models on serverless GPUs. Closed models through BYOK. The unified API. First-party guardrails and tools. Prepaid credits, a usage dashboard, and encrypted key storage.

  2. Phase 2Next

    Managed RAG as plugins. Shared GPUs for the most-used models. Custom Hugging Face model IDs. Spend caps.

  3. Phase 3Planned

    Open marketplace: plugin SDK, review process, publisher payouts. Harness templates. Fine-tuning presets.

  4. Phase 4Planned

    Invoiced billing. SOC 2. Private deployment. Team roles.

Request early access →