Certified Models by Hardware

Spectro Cloud certifies a small set of large language models (LLMs) for the hardware that PaletteAI Inference Launchpad supports. A certified model is one that Spectro Cloud has validated to run correctly on the listed GPU configuration. For what certification covers, refer to Model Certification.

info

This is not an exclusive list. You can bring your own model as long as it fits within the available GPU resources. Refer to Bring Your Own Model. If you want Spectro Cloud to certify a model for your hardware, contact Spectro Cloud.

NVIDIA

GPU configuration	GLM 5.2	DeepSeek v4 Pro	Kimi 2.7	Gemma 4
4 x H100	❌	❌	❌	✅
8 x H100	❌	❌	❌	✅
4 x H200	❌	❌	❌	✅
8 x H200	✅	✅	✅	✅
4 x B200	❌	❌	❌	✅
8 x B200	✅	✅	✅	✅

AMD

GPU configuration	GLM 5.2	DeepSeek v4 Pro	Kimi 2.7	Gemma 4
4 x MI300X	❌	❌	✅	✅
8 x MI300X	✅	❌	✅	✅
4 x MI325X	✅	❌	✅	✅
8 x MI325X	✅	❌	❌	✅
4 x MI350X	✅	✅	❌	✅
8 x MI350X	✅	✅	❌	✅
4 x MI355X	✅	✅	❌	✅
8 x MI355X	✅	✅	❌	✅

Vision Models

The table above lists certified text models. A text-only model cannot read screenshots or other images on its own. To answer questions about images, deploy a small vision model next to the text model and turn on vision preprocessing.

The certified pairing is GLM 5.2 as the text model and Qwen 3.5 9B multimodal as the vision model. Spectro Cloud has validated this pairing on the following configurations.

GPU configuration	Text model	Vision model	Validated
8 x MI325X	GLM 5.2	Qwen 3.5 9B multimodal	✅
8 x B200	GLM 5.2	Qwen 3.5 9B multimodal	✅
8 x H200	GLM 5.2	Qwen 3.5 9B multimodal	✅

On each configuration, the text model runs at tensor-parallel width 8 across all GPUs, and the vision model runs at tensor-parallel width 4 across the first 4 GPUs. Both models share the same physical devices; the appliance isolates their memory budgets so they do not compete for the same VRAM allocation.

On each shared GPU, the text model reserves approximately 86 percent of VRAM and the vision model reserves a further 4 percent, leaving roughly 10 percent headroom. The remaining memory is not spare capacity for an additional large model.

For how the request path works, refer to Vision Preprocessing. To deploy the pairing, refer to Enable Vision Preprocessing.

The tuned deploy configurations for both halves of the pairing live in the appliance model catalog as glm-5.2-shared and qwen-3.5-9B-shared. Each carries per-GPU-family variants for MI325X, B200, and H200, with the memory-budget, tensor-parallel, and engine-argument settings that were validated on the hardware above. Operators do not edit these directly; deploying the two catalog entries is enough.