Error: Failed to load model

Updated Jun 13, 2026 ·

Overview

LM Studio may fail to load or run a model with a generic error such as:

(Exit code: 18446744072635810000). Unknown error.

In some cases, the model actually loads successfully and the real issue only becomes visible in the logs.

For older NVIDIA GPUs, the root cause may be:

CUDA error: no kernel image is available for execution on the device

This usually means the GPU architecture is too old for the CUDA version or llama.cpp build being used.

Example Environment

Hardware used during troubleshooting:

Component	Specification
CPU	Dual Intel Xeon E5-2670 v3
Cores	24 Physical / 48 Logical
Memory	64 GB RAM
GPU	NVIDIA Quadro K2000
VRAM	2 GB

Troubleshooting Checklist

1. Look for generated tokens:

   n_decoded = ...

   If tokens are generated, the model loaded successfully.

2. Search the logs for:

   CUDA error

   or

   no kernel image is available for execution on the device

3. Disable GPU Acceleration

   GPU Offload = 0

   and retry.

4. Use a Smaller Model

   Llama-3.2-1B-Instruct

   or

   Qwen2.5-1.5B-Instruct

5. Reduce Parallelism

   Max Concurrent Predictions = 1

Investigation

When LM Studio fails to load a model, the first step is to check the logs for more detailed error messages.

LM Studio reports a generic model loading error:

(Exit code: 18446744072635810000). Unknown error.

Logs may contain:

LMSCore.getAvailableVRAMByDevice failed:
Could not find or load NVML library

and later:

CUDA error: no kernel image is available for execution on the device

The important thing to verify is whether the model actually loaded before the crash.

In the example below, the model successfully generated tokens:

n_decoded = 384
eval = 20.65 t/s

This indicates:

• The GGUF file is valid
• The model loaded correctly
• The CPU inference engine is working

The failure occurred afterward when CUDA attempted to execute on the GPU.

Root Cause

The NVIDIA Quadro K2000 uses an older CUDA architecture.

Modern CUDA builds used by the following tool may no longer support this GPU generation.

• LM Studio
• llama.cpp
• GGML CUDA backends

When CUDA attempts to launch kernels on unsupported hardware, the following error appears:

CUDA error: no kernel image is available for execution on the device

This means the GPU cannot run the CUDA code because it was compiled for newer architectures.

Recommended Solution

For older GPUs such as the Quadro K2000, the simplest and most reliable solution is to disable GPU acceleration and run models using CPU inference only. This avoids CUDA compatibility issues while still providing acceptable performance for smaller GGUF models.

Run the model using CPU only.

1. Disable GPU acceleration:

   GPU Offload: 0

2. If text is available:

   CUDA: Disabled
   Vulkan: Disabled
   GPU Acceleration: Disabled

Recommended LM Studio configuration

For a system with:

• 64 GB RAM
• Dual Xeon CPUs
• Older NVIDIA GPU

Start with:

CPU Thread Pool Size: 24
Eval Batch Size: 128
Physical Batch Size: 128
Max Concurrent Predictions: 1
Context Length: 2048
GPU Offload: 0

Parallelism Configuration

The logs showed:

n_parallel = 2

For CPU-only inference, start with:

n_parallel = 1

In LM Studio this typically means:

Max Concurrent Predictions: 1

This reduces resource contention and simplifies troubleshooting.

Example Small Models

When running entirely on CPU, smaller GGUF models are recommended.

Examples:

• Llama 3.2 1B Instruct
• Qwen 2.5 1.5B Instruct
• TinyLlama 1.1B Chat

Recommended quantizations:

Q4_K_M
Q4_0
Q5_K_M

A Q8 model can still run, but uses more RAM and may provide little practical benefit compared to Q4 or Q5 quantizations.

Additional: NVML Warning

You may also see:

Could not find or load NVML library

This warning indicates LM Studio cannot query NVIDIA GPU statistics.

It is usually not the primary cause of model loading failures.

The more important error is:

CUDA error: no kernel image is available for execution on the device