Granite Switch composes chat, decisions and speech in one model

Focus: Technical AI Engineering
Date: October 15, 2026
Topics: Activated LoRAs, Shared KV cache, Modular inference
Evidence: Publisher Authored
Availability: Available

An exploded chat base, decision and guardian adapters share KV context while a separate speech encoder produces transcription, with distinct decision and safety outputs and a workload-evaluation strip.

Summary: IBM Granite’s 8B and 30B Switch 4.1 checkpoints combine a general decision adapter, a guardian adapter and a speech encoder with the base chat model. Activated LoRAs reuse the base model’s KV cache, while the project supports vLLM and llama.cpp. Reported speed and benchmark claims remain publisher-produced and need workload-specific replication.

Why it matters: Evaluate the composed checkpoint as a system: verify adapter routing, classification quality, speech accuracy, cache behavior, memory use and latency on your own traffic. Treat publisher comparisons as hypotheses until identical fixtures reproduce them under the intended runtime.

Source: Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

How useful was this?

Anonymous feedback. No name or email collected.

← Daily Brief for October 19, 2026