VRAM stays about what the base model already used
IBM posted Granite Switch 4.1 at 8B and 30B on Hugging Face. The structured publish time is 2026-10-15 20:34 UTC, 4:34 a.m. in Beijing on October 16. One hosted endpoint can move among natural-language replies, general and specialized decisions, and speech to text.
The bundle is a general Decision adapter, a Guardian adapter for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC. Those three, plus the base model, take up approximately the same VRAM as the base model alone. The KV cache from a base-model generation can be reused for later adapter decisions.
[1]
The average win and the speed claim have no deltas here
They write that the Guardian adapter, on average, outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks. This post does not say by how much. They call Granite-Speech-5.0-Turbo-CTC the fastest ASR model on the OpenASR leaderboard. The leaderboard scores are not in this post.
In one 8B demo space, the same endpoint transcribes audio, writes summaries, screens for harmful content, and tests 15 other user-defined flags.
[1]The numeric rate belongs to a custom 3B
They compared Granite-Switch-4.1-30B with Jev-27B. The first is a frozen Granite base plus four specialized adapters. The second is a frozen Qwen base plus one generic decision adapter. The 30B finished a real-world five-step retrieval workflow much faster than its competitor. The post does not say how much faster.
The number is a different setup. A custom 3B Granite Switch made 60 to 100 decisions a second while navigating Doom, against the 10 decisions a second Jev reported. That is not an 8B or 30B result. The 3B can also answer questions while it runs. An adapter SDK is something they plan to release in Mellea. More detail and data are described as coming in the weeks ahead.
[1]