An open-weight model, one whose trained files (the weights) anyone can download, goes live only after its weights are published. We start work on a new open-weight model as soon as its weights are released. We shipped Kimi K3 the day after Moonshot released them.
Plan a switch around the weights date and our model list, not the launch post. Kimi K3.1 Pro shows the gap. Moonshot's announcement of 24 October 2026 made it available in Kimi's own apps and API, with the full weights due by 4 November.
So when can you actually call a new model such as Kimi K3.1 Pro on a European API? Engineering teams ask us this in sales calls, usually right after a big release. One customer put it this way: "Do you have a zero-day implementation of the weights for new model releases such as Kimi K 3.1 Pro?"
Day zero starts when the weights are published
A model launch usually has three dates, and only one of them starts the clock for an open-model provider. The vendor announces the model. The vendor often serves it first on its own API. Then the vendor publishes the weights.
Kimi K3.1 Pro follows this pattern. TNW reported that Kimi K3.1 Pro is available in Kimi's apps and API, and that Moonshot plans to publish the full weights by 4 November. Until those files are public, no provider outside Moonshot can serve the model.
Lyceum is a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland, so a European team can call open models on European GPUs. Prompts and outputs are processed, not stored, and never used for training. That only works once the weights exist, which is why the weights date matters more to you than the launch post.
What happens between the weights and the first API call
Between the weights and your first request sit two jobs. First, the serving engine has to support the model. A serving engine is the software that loads the weights onto GPUs and answers requests, such as vLLM.
New models often bring new architecture parts, and the engine has to learn each one. Some engine teams are ready on release day. The vLLM project announced that it supported the TML Inkling model "on Day 0", the day the model was released.
Second, a provider has to run the model on its own GPUs. It loads the weights and starts replicas, identical copies of the model that share the traffic. Then it tests the outputs and publishes a model ID and a price. The model ID is the name you put in the model field of your request.
For Kimi K3, both jobs took us one day. Our newsletter of 28 July 2026 said we shipped it the day after Moonshot released the weights. That is one precedent, not a promise for every model: we prioritise new models by model quality and customer demand.
Why a new model keeps changing in its first days
A model that answers on day one is not finished. Engines keep improving how they run a new architecture after launch. When Llama 4 came out, the vLLM team supported it at release and wrote that "more performance enhancements are on the way".
Behaviour can also differ from model to model, even well after launch. A tool call is a request in which the model asks your code to run a function, such as a search. On 17 September 2026, we checked how models handle a request that requires a tool call. deepseek/deepseek-v4.1-flash and z-ai/glm-5.2 did not always call the tool.
Vendors also ship updated versions of a model after launch. When DeepSeek released V4.1 Flash 1017, we kept the earlier DeepSeek V4.1 Flash for a one-week grace period until 27 October 2026. After that, requests to the old model ID went to the new version automatically. Callers did not have to change anything.
So the first days of a model are a test period for you as well. Run the features your product relies on, such as tool calls and output formats, before you trust the model with production traffic.
Kimi K3.1 Pro: the launch timeline worked through
Kimi K3.1 Pro is not live on our API yet. We haven’t run it, so our Kimi K3.1 Pro model page shows no figures of ours. It lists the model as "Coming soon to Lyceum". The price, model ID and region appear there once it goes live.
The table puts the known dates next to a calculated one. The last row is hypothetical.
| Step | Date | Source |
|---|---|---|
| Moonshot's announcement published | 24 October 2026 | our Kimi K3.1 Pro page |
| Full weights due | by 4 November 2026 | TNW |
| Earliest call on our API at the Kimi K3 pace | 5 November 2026 (hypothetical) | calculated below |
The calculation uses only those dates and the one-day gap we needed for Kimi K3.
announcement to latest weights date: 24 Oct to 4 Nov = 11 days
weights to first call (Kimi K3 pace): +1 day
announcement to earliest call: 11 + 1 = 12 days (5 Nov, hypothetical)In this example, 12 days pass between the launch post and a possible first call. Eleven of them are the wait for the weights. The example favours an early date: it assumes the weights land on 4 November and that our Kimi K3 pace repeats. It says nothing about the model's speed, quality or price. If Moonshot publishes the weights earlier, every later date moves earlier too.
How to check that a model is live on our API
The model list of our API is the one place that tells you whether a model is callable right now. The GET /models endpoint is the authoritative list of IDs our serverless API accepts, and it reflects what is served at that moment. The base URL is https://api.lyceum.technology/openai/v1.
This request lists every model ID our serverless API accepts and filters the output for Kimi models:
curl -s https://api.lyceum.technology/openai/v1/models \
-H "Authorization: Bearer $LYCEUM_API_KEY" | grep -i kimiA model can appear in the list while its backend is briefly unavailable. Our API then answers 404 model not found, the same as for an unknown ID. Poll the list or retry with backoff before you treat that answer as final.
The models page shows each live model with its price and its context window (the most text the model reads in one request). For incidents on models that are already live, check our status page.
Run a new model early on a dedicated deployment
A dedicated deployment runs one model on GPU replicas reserved for you, behind its own OpenAI-compatible endpoint. Hugging Face is the public hub where vendors usually publish open weights. Our dedicated inference docs state that any Hugging Face model ID can be passed to POST /inference/create.
That gives you a second route once the weights are public. You do not have to wait for a model to appear in our serverless list. You can deploy it yourself, provided the serving engine already supports its architecture.
The trade-off is cost. A dedicated deployment is billed for the time its replicas run, not per request. It pays off for a team that needs the new model early and has steady traffic to keep the replicas busy. For a short evaluation, the serverless model is the cheaper route once it is listed.
When to move production traffic to a new model
Move production traffic only when three conditions hold. The model ID appears in our model list. Your own tests pass on the features you depend on. And the model has run through a trial on a sample of your real traffic.
Keep your current model ID in your configuration as a fallback until then. Switching back is one line in your request, so a failed test does not have to become an outage. Compare the two models on your own prompts, not on the vendor's benchmark table.
If you are waiting for Kimi K3.1 Pro, start with Kimi K3 now and test your workload on it. Open the Kimi K3.1 Pro page to see when it goes live, and run Kimi K3 today.
Frequently asked questions
How quickly does Lyceum add newly released open-weight models, and can customers request a specific model?
As fast as we can. Our team starts on new open-weight models as soon as the weights are released, and we prioritise based on model quality and customer demand. If you need a specific model, tell us and we'll look at adding it.