Supported Models

Paddock implements most common and newer model families natively rather than a huge library of models.

How Support Works

Every model family in Paddock has its own native tuning in the engine, running CUDA kernels. Each release contains a built-in catalog of models, downloaded from our fast object storage.

Generative Models

ModelArchitectureModalitiesWeights
GPT-OSS 20BMixture of Experts (32 experts)TextMXFP4
GPT-OSS 120BMixture of Experts (128 experts)TextMXFP4
Qwen3.5 9BDense hybrid (Gated DeltaNet + gated attention)TextQ8_0
Qwen3.6 27BDense hybridText + imagesQ8_0
Qwen3.6 35B A3BHybrid Mixture of Experts (256 experts)Text + imagesQ8_0
Gemma 4 31BDense (interleaved sliding-window and global attention)Text + imagesQ8_0

Notes on individual families:

  • GPT-OSS models use the Harmony output format; the server parses reasoning and tool calls natively, and reasoning effort (low, medium, high) is exposed through the API.
  • Qwen3.5 and Qwen3.6 share one hybrid architecture in the engine: three Gated DeltaNet linear-attention layers for every full-attention layer. The dense and MoE variants load through the same code path, keyed by GGUF metadata.
  • Qwen3.6 35B A3B ships a built-in multi-token-prediction (MTP) head in its GGUF, which the engine uses for speculative decoding.
  • Gemma 4 31B is the newest family in the engine. Its speculative-decoding drafter is a separate small GGUF loaded with the --mtp flag rather than a built-in head.

Embedding and Reranker Models

Encoder models are served through POST /v1/embeddings and POST /v1/rerank instead of the chat routes. The server detects an encoder architecture at load time and routes accordingly.

ModelTaskWeights
Qwen3 Embedding 0.6B / 4B / 8BText embeddings (Matryoshka dimensions)Q8_0
Qwen3 Reranker 0.6B / 4B / 8BQuery-document relevance scoringQ8_0

At load time the server runs a short, deterministic quality calibration for encoder models to pick the fastest kernel profile that holds retrieval quality for that specific model. The verdict is cached on disk, so it costs a few seconds once per model and pack. It can be disabled with PADDOCK_BS_CALIB=off.

Vision Input

Qwen3.6 27B, Qwen3.6 35B A3B, and Gemma 4 accept image content parts in chat requests.

When a pdfium library is configured, the server also accepts PDF attachments on the vision path: pages are rasterized server-side into images, with a page cap. See Server Configuration for the PDF settings.

GGUF Loading and Quantization

Paddock loads standard GGUF files, memory-mapped from disk. The formats the engine executes today:

FormatStatusUsed by
Q8_0Executed on GPUQwen3.5 / Qwen3.6 / Gemma 4 / Qwen3 encoders
MXFP4Executed on GPUGPT-OSS expert weights
F32 / F16 / BF16Executed on GPUNorms, embeddings, vision towers

For measured serving throughput per model, see the benchmarks.