Skip to main content
Pioneer supports two model families: encoder models (GLiNER) for structured extraction tasks like named entity recognition, and decoder models (LLMs) for text generation, classification, and open-ended prompting. The tables below are a snapshot of the current catalog — use GET /base-models to query the live list, which always reflects current availability and capabilities. Some rollout-stage models are feature-gated. They appear in the live catalog only for workspaces that have the corresponding rollout enabled.

​
Encoder models (GLiNER)

GLiNER models perform named entity recognition and structured extraction. Most GLiNER base models support both training and on-demand inference after training. Prices are per 1M tokens. fastino/gliner2-multi-v1 and fastino/gliner2-multi-large-v1 are multilingual variants suitable for non-English text.

​
Decoder models — training

LoRA fine-tuning via POST /felix/training-jobs is limited to the Nemotron 3.5 Lightning family: Together with the trainable GLiNER2 Base, Large, Multi, and Multi Large encoder targets above, these are the only supported base models for new training jobs. Query GET /base-models?supports_training=true before submitting a job; it is the live source of truth for availability.

​
Decoder models — serverless inference

These are the supported pre-deployed inference families. Query GET /base-models?supports_inference=true for live availability, context limits, and pricing.

​
Nemotron 3.5 Lightning

​
Anthropic

​
OpenAI

​
DeepSeek

​
Z.ai

Models outside these families are not supported inference targets. Use GET /base-models for the live catalog and lifecycle state before integrating.

​
Prompt caching

Many serverless models bill cached input tokens at a discount, and some providers bill a one-time surcharge to write tokens into the cache. Pioneer passes each provider’s published cache rates straight through — these are the same rates GET /base-models returns as cache_read_price_per_million and cache_write_price_per_million, and the same rates Pioneer bills you. Cache rates are derived from each model’s input rate using the multipliers below. Where a provider has no separate cache-write line item, cache writes bill at the standard input rate. Query the live catalog for exact cache rates. Any model without an explicit cache discount bills cached input at the standard input rate.

​
On-demand vs. serverless inference

Pioneer offers two ways to serve predictions, and the right choice depends on your workflow. Serverless inference uses pre-deployed base model endpoints. There is no startup delay and you are billed per token. This is ideal when you want to call a frontier model without fine-tuning. On-demand inference provisions a dedicated GPU after fine-tuning completes. Your LoRA adapter is loaded onto the GPU and served exclusively for your requests. Pioneer routes inference calls to an on-demand deployment automatically when you pass a training job ID as model_id.

​
Querying the live catalog

The tables above may lag behind newly added models. Use GET /base-models to get the current catalog at runtime.
Each entry in the response includes the model ID, its display label, context length, per-1M-token rates (input_price_per_million, output_price_per_million, cache_read_price_per_million, cache_write_price_per_million), and boolean flags for supports_training and supports_inference. Use the model ID value directly in training job requests and inference calls. To list every model alongside its input, output, and cache rates: