How to use open source models with Langdock

Langdock is a good fit for companies that want a ChatGPT-style workspace with SSO, agents and workflows, without building it themselves. The catch is the model list. Out of the box you get the big closed labs. If your team wants DeepSeek for cheap long-context work, GLM or Kimi for coding, or an open-weight model because your data policy says weights must be inspectable, you have to bring it yourself.

Langdock lets you do that through an OpenAI Compatible key. Point that key at Requesty and every open-weight model in our catalog becomes available to your workspace: DeepSeek V4.1 Flash, GLM-5.3 and GLM-5.3 Flash, Kimi K3, Nemotron 3 Ultra, MiMo V2.5 and a few hundred more. One key, one base URL, one invoice.

This post walks through the setup and then covers the questions that come up right after: which models to start with, how to keep traffic in the EU, and how to see who is spending what.

Why route through a gateway instead of a single provider

You could paste a DeepInfra, Fireworks or Novita key straight into Langdock. Three things push most teams to a gateway instead:

1. Open-weight models are served by many providers, with very different prices and speeds. The same DeepSeek V4.1 Flash weights are hosted by seven providers in our catalog. Requesty picks the endpoint per request based on your routing policy (cheapest, fastest, EU-only) and fails over when one is down.
2. Langdock needs one key per provider. Every provider you add is another credential to rotate and another vendor to onboard. A gateway collapses that to one.
3. You want to see usage per user and per model. Langdock shows you chat history. Requesty shows you cost, tokens, latency and cache hit rate per model and per key, which is what finance asks for.

Step 1: create a Requesty API key

Sign in at app.requesty.ai, open API Keys, and create a key named Langdock. If you want to restrict which models this key can call, set up an approved-model list first. That is also how you enforce "open-weight only" or "EU-hosted only" at the gateway level rather than trusting each user to pick the right model.

Step 2: add the key in Langdock

In Langdock go to Settings > Models, scroll to Keys, and click Add key. Choose OpenAI Compatible and fill in:

• Key name: Requesty
• Base URL: https://router.requesty.ai/v1, or https://router.eu.requesty.ai/v1 for EU-only routing
• API Key: the key from step 1
• Hosting Provider: Requesty

Save. The key shows as Not used until a model references it.

Step 3: add each open source model as a custom model

Langdock does not pull the model list from the provider, so each model is added by hand. Click Add custom model, ignore the pre-configured list, and click set up manually in the bottom right.

On the Configure step:

• Type: Completion
• Model Name: what users will see, for example DeepSeek v4.1 Flash
• Provider Model Name: the Requesty model ID, for example deepseek-v4.1-flash
• Provider: Other
• Region: Global, or EU if you used the EU base URL

On the Deployment step, pick the Requesty key, paste the same model ID, and click Test & continue. Langdock sends a real request through Requesty and only lets you save once it succeeds.

The bare name deepseek-v4.1-flash is a Requesty managed policy. Instead of pinning one provider, Requesty picks the best endpoint for that model on every request and fails over when a provider is down. If you want a specific provider, use the provider/model ID from the catalog instead, for example fireworks/deepseek-v4.1-flash.

Repeat for every model you want to expose. Each one takes about a minute because the key is already there.

Step 4: test it in chat

Open a new chat, pick the model from the selector, and ask it what it is.

The request shows up in your Requesty logs with the provider that served it, the tokens used and the cost.

Which open source models to add first

There is no single best open-weight model, but our traffic points to a clear shortlist. These are the open-weight models with the most traffic routed through Requesty right now, ranked by token volume, with the blended price customers paid after caching:

The ranking comes from our public model economics data. DeepSeek V4.1 Flash went from zero to a fifth of all our traffic in under a month, so this list will change. Each ID is a managed policy, so Requesty maintains the provider chain behind it. Check the model library for the current list, prices and which providers host each model.

Keeping traffic in the EU

Most open-weight models are hosted by at least one EU provider, and several are served from Frankfurt or Paris. Two settings make the whole Langdock workspace EU-only:

1. Use https://router.eu.requesty.ai/v1 as the base URL in step 2.
2. Use the @eu variant of each managed policy as the model ID, for example deepseek-v4.1-flash@eu or kimi-k3@eu, and set Region to EU on the custom model.

With the EU router and @eu policies, Requesty only considers EU-hosted endpoints for the model you chose, and requests are processed in Frankfurt. Combine that with an approved-model list that only contains EU policies and users cannot pick a non-compliant model by accident.

Tracking cost per model

Every request through the Langdock key is logged in Requesty with model, provider, tokens, cost, latency and cache status. The analytics dashboard groups spend by model and by key, so you can see whether the switch to open-weight models paid off, and which models people use. Cache hit rates matter here: on our routed traffic, DeepSeek V4 Flash 0731 saw an 81% cache hit rate and GLM-5.3 Flash 93%, which cut bills by more than half compared to list price.

Common issues

• Test & continue fails: the model ID does not match a catalog entry, or the key is not allowed to use it. Copy the ID from the model library and check the approved-model list.
• 401 Unauthorized: the key was pasted with a trailing space or has been revoked. Edit the key in Langdock and paste it again.
• Model missing in chat: make sure Visible to users is on and the deployment toggle is enabled.

Wrapping up

Langdock gives you the workspace. Requesty gives you the models. Add one key, add the open-weight models your team wants, and you have DeepSeek, GLM, Kimi or Nemotron next to GPT and Claude in the same chat selector, with EU routing and per-model cost tracking already in place.

The full step-by-step reference with screenshots for each step lives in our Langdock integration guide.