Implementation:Ollama Ollama XCreate

Knowledge Sources	Ollama
Domains	Model Creation, Safetensors
Last Updated	2025-02-15 00:00 GMT

Overview

Core model creation logic for the experimental safetensors format, providing model type detection, manifest management, and the creation pipeline for both LLM and image generation models.

Description

Resolves model names to manifest paths using the registry.ollama.ai convention. Loads and parses manifests and model configs to detect model types (safetensors LLM vs. image gen) via capabilities. CreateSafetensorsModel iterates safetensors files, extracts individual tensors, optionally quantizes them, and creates blob layers via callback functions (LayerCreator, QuantizingTensorLayerCreator, ManifestWriter) to avoid import cycles. Supports architecture detection for parser/renderer selection, thinking capability detection, and expert group packing for MoE models.

Usage

Central package for the experimental safetensors model format, used by both client-side and server-side creation paths.

Code Reference

Source Location

• Repository: Ollama
• File: x/create/create.go
• Lines: 1-588

Signature

type ModelConfig struct {
    ModelFormat  string   `json:"model_format"`
    Capabilities []string `json:"capabilities"`
}

type LayerInfo struct {
    Digest    string
    Size      int64
    MediaType string
    Name      string
}

type LayerCreator func(r io.Reader, mediaType, name string) (LayerInfo, error)
type QuantizingTensorLayerCreator func(r io.Reader, name, dtype string, shape []int32, quantize string) ([]LayerInfo, error)
type ManifestWriter func(layers []LayerInfo) error

func IsSafetensorsModel(modelName string) bool
func IsSafetensorsLLMModel(modelName string) bool
func IsImageGenModel(modelName string) bool
func IsSafetensorsModelDir(dir string) bool
func CreateSafetensorsModel(modelName, modelDir, quantize string, ...) error

Import

import "github.com/ollama/ollama/x/create"

I/O Contract

Inputs

Name	Type	Required	Description
modelName	string	Yes	Target model name (e.g. "my-model:latest")
modelDir	string	Yes	Path to directory containing safetensors files
quantize	string	No	Quantization type: "int4", "int8", "nvfp4", or "mxfp8"

Outputs

Name	Type	Description
error	error	Non-nil if model creation fails

Usage Examples

// Check if a model uses safetensors format
if create.IsSafetensorsLLMModel("my-model:latest") {
    // Use MLX runner instead of llama.cpp
}

// Detect model directory type
if create.IsSafetensorsModelDir("/path/to/model") {
    // Has config.json + .safetensors files
}

Related Pages

• Principle:Ollama_Ollama_ModelCreation