CIQ Fuzzball 4.3 adds preset deployment for open-weight models on private GPUs

CIQ announced Fuzzball 4.3 on October 8, adding a catalog-based route for organizations that want to serve open-weight models on GPUs they control. This is infrastructure news, not a new foundation-model checkpoint. CIQ says users can start supported models from presets and connect clients or agents through an OpenAI-compatible endpoint. The company also documents a general vLLM catalog entry for compatible Hugging Face models, designed for deployment across NVIDIA and AMD GPU infrastructure.

The release lists eleven validated presets: GPT-OSS 20B and 120B, Ministral 3 8B and 14B, Qwen3-Coder 30B and Next 80B, Nemotron 3 Nano and Super, Mistral Medium 3.5, Llama 4 Scout and Gemma 4 31B. In CIQ's terminology, presets supply selected GPU counts, context sizes and serving arguments. That is a vendor product-validation claim, not evidence that Open Model Weights has independently run every model or measured its latency, accuracy or memory usage.

Fuzzball 4.3 adds replica pools behind stable endpoints. CIQ describes automatic scale-up and scale-down, including scaling idle workloads to zero, and LiteLLM gateways for discovery and access control. The company also highlights agents that can find model endpoints within a cluster. These features may be useful for private inference deployments, but the release is not a benchmark report and does not establish a measured cost-per-token advantage over alternative infrastructure.

CIQ documents operational constraints that should accompany any future compatibility claim. Local persistent volumes cannot support concurrent multi-node stages, and autoscaled serving needs shared storage when replicas span nodes. The upgrade also changes catalog identifiers and tightens HTTPS requirements for catalog sources. Open Model Weights should retain those conditions with source-linked runtime evidence, rather than converting catalog support into a blanket compatibility guarantee. No model's weights, license or measured performance changed merely because this orchestration software was released.