IBM says Granite Switch 4.1 fits chat, safety screening and speech in one base model's VRAM

We think the weights argument is sound. In IBM's own race setup the base Granite is frozen and the specialists are four aLoRAs on top of it ^[15]. Each specialist is a set of adapter weights on a shared base. IBM draws the conclusion in the post: "That means more capabilities with fewer GPUs." ^[11]

The GPU count is a separate question. The text of the post does not include a measured memory figure for the 8B or the 30B, and "approximately the same" is the whole claim ^[10]. RuntimeWire lists the Hugging Face post as its primary source ^[6], so its piece is a second reading of one document, not a second measurement. It does name the open question: whether the savings hold in production "remains the central test for developers considering a composed model over separate services" ^[5].

The cache design is where the larger saving sits. A standalone guardian model has to read the conversation itself. A Switch adapter starts from the cache the base model already built ^[14]. That removes a second prefill and a second copy of the cache for the same text. IBM also says that adding specialists next to the generalist "costs almost nothing" ^[16].

The benchmark claims are IBM's, on IBM's choice of comparison. The Guardian is said to beat four named decision models on average ^[12]. That transfers only if your safety categories look like the tasks IBM averaged over. The speech model is described the same way, as the fastest on the OpenASR Leaderboard ^[13]. IBM also says its specialized adapters can beat generalists on narrow classification tasks ^[17], which is a point for custom composition and a caution about any single average.

The race changes the base model and the adapter design together. It puts a 30B Granite with four adapters against Jev-27B, a Qwen base with one generic decision LoRA ^[15]. IBM says Granite finished a 5-step RAG workflow "much faster". A speed gap alone cannot say which change produced it.

The Doom figure is a 6x to 10x gap ^[9]. That is 60 to 100 decisions a second for a custom 3B Granite Switch against 10 reported for Jev ^[4]. The 3B model is a demo build. The released models are the 8B and 30B ^[7]. The game loop is a fair test of decisions per second and a poor one of guardrail latency. Jev's 10 is Jev's own figure, cited in the post to a Typesafe blog ^[8].

IBM's demo space runs the 8B on one endpoint for transcription, summaries, harmful-content screening and 15 other user-defined flags ^[1]. The measurement that would settle the GPU question is peak memory with every adapter loaded at a team's own concurrency, set against the sum of the servers it runs now.