AI apps

Blender 5.2.2 for headless Cycles rendering over SSH, with the build that fits the GPU: CUDA (NVIDIA), HIP (AMD) or oneAPI (Intel Arc).

ComfyUI v0.37.4, node-based image and video generation; bring your own models; the same template runs on NVIDIA (CUDA), AMD (ROCm) and Intel Arc (XPU) GPUs.

ComfyUI that downloads FLUX.1 [schnell] (fp8, Apache-2.0) at start.

ComfyUI that downloads SDXL base 1.0 at start; fits 8 GB cards.

NVIDIA CUDA 12.4 + cuDNN development image with SSH.

NVIDIA CUDA 12.9 + cuDNN development image with SSH.

NVIDIA CUDA 13.4 + cuDNN development image with SSH.

DeepSeek-R1 reasoning distilled into Llama 8B, on llama.cpp.

• Qwen 1.5B
• Qwen 7B
• Llama 8B
• Qwen 14B
• Qwen 32B · vLLM
• Qwen 32B · llama.cpp

Mistral's coding agent model, 4-bit GGUF on llama.cpp.

Dolphin 3.0 on Llama 3.1 8B, an uncensored fine-tune, 4-bit GGUF.

Google's Gemma 3 12B, 4-bit GGUF, on llama.cpp; fits 12 GB cards; runs on NVIDIA, AMD, Intel Arc and Apple silicon.

• 270M
• 4B
• 12B
• 27B

Google's on-device Gemma 3n, 8-bit GGUF on llama.cpp.

Google's Gemma 4 12B, 4-bit GGUF on llama.cpp.

• 12B
• 26B A4B
• 31B

GLM-4 32B 0414, 4-bit GGUF on llama.cpp.

OpenAI's open-weight 120B reasoning model on vLLM; needs an 80 GB card.

• 20b · llama.cpp
• 20b · vLLM
• 120b

OpenAI's open-weight 120B reasoning model on llama.cpp.

Mistral Small 24B with refusals removed (abliterated), 4-bit GGUF.

Qwen3 14B with refusals removed (abliterated), 4-bit GGUF.

QwQ 32B with refusals removed (abliterated), 4-bit GGUF.

SciPy notebook with JAX 0.11.2 and its CUDA 12 wheels.

Jupyter Docker Stacks PyTorch notebook with a per-rental token; the same template runs on NVIDIA (CUDA) and AMD (ROCm) GPUs.

Jupyter Docker Stacks TensorFlow notebook with a per-rental token.

Kokoro-82M text-to-speech with an OpenAI-compatible /v1/audio/speech API.

Meta's Llama 3.1 8B Instruct, 4-bit GGUF on llama.cpp.

Meta's Llama 3.2 1B, a small chat model on llama.cpp.

• 1B
• 3B

Meta's Llama 3.3 70B, AWQ 4-bit, on vLLM; needs an 80 GB card.

• 70B · vLLM
• 70B · llama.cpp

LLaMA-Factory's web UI and CLI for fine-tuning 100+ model families.

llama.cpp CLI tools, converters and quantizers for making your own GGUF files.

EleutherAI lm-evaluation-harness 0.4.13 with vLLM for fast benchmark runs.

Mistral's reasoning model, 4-bit GGUF on llama.cpp.

Mistral 7B Instruct v0.3, 4-bit GGUF on llama.cpp.

Mistral Small 3.2 (24B), 4-bit GGUF with a 16K context, on llama.cpp; fits 24 GB cards; runs on NVIDIA, AMD, Intel Arc and Apple silicon.

Text embeddings for search and RAG on llama.cpp.

Ollama daemon; pull models on demand and call its API through an SSH tunnel; the same template runs on NVIDIA (CUDA) and AMD (ROCm) GPUs.

Ai2 OLMo 2 7B Instruct, a fully open model, 4-bit GGUF.

Ai2 Olmo 3 7B Instruct, a fully open model, 4-bit GGUF.

Ai2 Olmo 3.1 32B Instruct, a fully open model, 4-bit GGUF.

Ai2 Olmo 3.1 32B Think, a fully open reasoning model, 4-bit GGUF.

Ai2 OLMoE 1B-active 7B mixture of experts, 4-bit GGUF; fits 8 GB laptops.

Chat UI with a bundled Ollama: models, documents, tools and agents in the browser.

Microsoft Phi-3.5 mini 3.8B, 4-bit GGUF; fits 8 GB laptops.

Microsoft's Phi-4 14B, 4-bit GGUF on llama.cpp.

Microsoft's Phi-4 mini, a small chat model on llama.cpp.

Microsoft Phi-4 reasoning 14B, 4-bit GGUF on llama.cpp.

PyTorch 2.14 on CUDA 13.0 for Blackwell (RTX 50) and newer cards; the same template runs on NVIDIA (CUDA), AMD (ROCm) and Intel Arc (XPU) GPUs.

Qwen2.5 14B Instruct, 4-bit GGUF on llama.cpp.

• 7B · TGI
• 7B · llama.cpp
• 14B
• 32B

Qwen2.5 Coder 1.5B with refusals removed (abliterated); fits any 8 GB machine.

Qwen2.5 Coder 1.5B, a small code model that fits any 8 GB machine.

Qwen2.5 Coder 14B for code completion and chat on llama.cpp.

• 14B
• 32B

Qwen3 1.7B, 8-bit GGUF on llama.cpp.

• 1.7B
• 4B
• 8B · SGLang
• 8B · vLLM
• 8B · vLLM
• 14B · vLLM
• 14B · llama.cpp
• 30B A3B
• 32B · vLLM
• 32B · llama.cpp

Qwen3 30B A3B Instruct 2507, mixture of experts, 4-bit GGUF.

• 4B
• 30B A3B

Qwen3 4B Thinking 2507, a small reasoning model in 4-bit GGUF.

Qwen3 Coder for editors and agents, 4-bit GGUF on llama.cpp.

Qwen3 text embeddings for search and RAG on llama.cpp.

Qwen3 Next 80B A3B Instruct, hybrid mixture of experts, 4-bit GGUF; needs an 80 GB card.

Qwen3.8 27B, an uncensored fine-tune, 4-bit GGUF.

QwQ 32B, Qwen's reasoning model, 4-bit GGUF on llama.cpp.

cuDF, cuML and cuGraph in JupyterLab (RAPIDS 26.08).

ByteDance Seed-OSS 36B Instruct, 4-bit GGUF on llama.cpp.

SSH workspace with SGLang installed for serving and benchmarking.

SmolLM2 1.7B Instruct, 8-bit GGUF; runs on small GPUs and 8 GB Macs.

SmolLM2 1.7B, an uncensored fine-tune, 4-bit GGUF; runs on small GPUs.

Hugging Face's SmolLM3 3B, 4-bit GGUF on llama.cpp.

faster-whisper speech-to-text (and Kokoro/Piper text-to-speech) behind an OpenAI-compatible API.

PyTorch-native fine-tuning recipes (torchtune 0.6.1) on PyTorch 2.7.

Unsloth's fast LoRA/QLoRA fine-tuning in JupyterLab.

SSH workspace with vLLM installed: serve any model you choose with `vllm serve`.

code-server 4.138 on PyTorch 2.8 / CUDA 12.8 devel, with a per-rental password.

WhisperX 3.8.6: fast transcription with word timestamps and diarization.

01.AI Yi 1.5 9B Chat, 4-bit GGUF on llama.cpp.