NeoCloud Test Results
The same kinds of engines and models we run on our own workstations, measured on rented GPUs from neocloud providers, starting with Google Colab. Each neocloud provider gets its own section as we test it, and every notebook is published on GitHub.Engines and models measured on neocloud GPUs, starting with Google Colab.
Google Colab
Our open notebooks run language and image models on Google Colab Pro GPU runtimes. Each one installs a single engine, downloads its model once to Google Drive and serves it through one authenticated OpenAI-compatible API. Measured on real Colab runtimes in November 2026, one request at a time.Our open notebooks on Google Colab Pro GPU runtimes, measured in November 2026.
Decode is a 512-token reply at temperature 0. Prefill speed and time to first token use a fresh prompt of about 3,500 tokens on a loaded model (no prefix-cache reuse); notebooks not yet rerun with it read “next test”. Image and video speed is per output after the first.
L4 · 24 GB (RTX 4070 Ti class)

| Model notebook | Engine | Prefill t/s | Time to first token | Decode t/s |
|---|---|---|---|---|
| nemotron-3.5-bolt-74b-iq2 | llama.cpp | —next test | —next test | 38 |
| qwen3.8-27b-w4a16-sglang | SGLang | —next test | —next test | 16 |
| flash-next-gyro-s | Strata | —next test | —next test | 68 |
| qwen3.8-27b-exl3-2bpw | TensorFold | —next test | —next test | 83 |
| Model notebook | Engine | Output | Time per output |
|---|---|---|---|
| flux3-klein-comfy | ComfyUI | 1024×1024 image | 8 s |
| minimax-h3-l4-comfy | ComfyUI | 832×480 5 s video clip | 124 s |
| qwen-image-2.1-turbo-comfy | ComfyUI | 1024×1024 image | 8 s |
| qwen-image-2.1-nf4 | Diffusers | Image | 37 s |
| sdxl-base | Diffusers | Image | 17 s |
| sdxl-turbo | Diffusers | Image | 3 s |
A100 · 80 GB (High-RAM)

| Model notebook | Engine | Prefill t/s | Time to first token | Decode t/s |
|---|---|---|---|---|
| nemotron-3.5-bolt-74b-iq2 | llama.cpp | —next test | —next test | 79 |
| sage-a1.5-q4 | llama.cpp | 3,558 | 1.00 s | 134 |
| qwen3.8-27b-w4a16-sglang | SGLang | 2,018 | 1.76 s | 74 |
| flash-next-strata-iq3-s | Strata | —next test | —next test | 167 |
| qwen3.8-27b-exl3-2bpw | TensorFold | —next test | —next test | 144 |
| Model notebook | Engine | Output | Time per output |
|---|---|---|---|
| flux3-klein-comfy | ComfyUI | Image | 3 s |
| minimax-h3-comfy | ComfyUI | 832×480 5 s video clip | 35 s |
| qwen-image-2.1-bf16-comfy | ComfyUI | 1024×1024 image | 9 s |
| qwen-image-2.1-turbo-bf16-comfy | ComfyUI | 1024×1024 image | 4 s |
| wan22-5b-comfy | ComfyUI | 832×480 5 s video clip | 61 s |
G4 · RTX PRO 6000 Blackwell, 96 GB

| Model notebook | Engine | Prefill t/s | Time to first token | Decode t/s |
|---|---|---|---|---|
| basalt-flash-next-nvfp4-dq6k | Basalt | —next test | —next test | 273 |
| hyperqwen-qwen3.8-27b-dflash2 | HyperQwen | —next test | —next test | ~315 |
| glm-maya-s | Maya | —next test | —next test | 77 |
| qwen3.8-27b-w4a16-sglang | SGLang | —next test | —next test | 77 |
The remaining A100 and G4 notebooks are still being tested.
Ready to get started?
Tell us about your environment, we’ll design a deployment that keeps your data yours.
