IBM Granite Switch 4.1 combines chat, decisions and English speech transcription

IBM’s Apache-2.0 8B and 30B checkpoints bundle chat, on-demand decision and safety adapters, and English-only ASR in one self-hostable model interface.

2 min read

Source code on a monitor, illustrating AI-assisted reverse engineering with Ghidra MCP.

IBM has released Granite Switch 4.1, offering 8B- and 30B-parameter configurations that combine chat, structured decisions, safety screening and speech transcription in one model checkpoint. The October 15 announcement frames the release as a modular deployment update: adapters activate on demand, rather than requiring a separate model endpoint for every capability.

The 8B model card lists a 128,000-token context window, a Decision Activated LoRA (aLoRA), a Guardian safety adapter and an embedded 470-million-parameter Granite Speech encoder. The 30B version uses the same configuration on a larger base. Both cards specify Apache 2.0 and support for Hugging Face Transformers and vLLM; weights are available from Hugging Face.

Granite Switch’s public GitHub repository, maintained by IBM’s generative-computing team, contains the composer and inference code for building and deploying these modular checkpoints. It showed 100 stars and 14 forks on October 16, 2026. That is a total-star snapshot, not evidence of recent star growth. The repository lists Apache-2.0.

The added speech path has important limits: it is English-only, returns lowercase, unpunctuated transcripts, and the encoder and language model are not jointly trained. IBM calls the speech component the fastest on the OpenASR leaderboard, but that is the vendor’s characterization, not an independent evaluation of the combined Granite Switch system. The release also points to IBM-reported decision benchmarks; those should be read as company results rather than third-party validation.

For teams testing a single self-hosted endpoint across chat, decisions and transcription, the release makes that configuration downloadable today. The main trade-off is that the combined interface does not remove component-specific constraints—especially the speech encoder’s language and formatting limits.

Source code on a monitor, illustrating AI model tooling.

Illustrative image, not a Granite Switch screenshot. Photo by Chris Ried on Unsplash, under the Unsplash License.

Sources: IBM Granite Switch 4.1 announcement · Granite Switch 4.1 8B model card · Granite Switch 4.1 30B model card · Canonical GitHub repository · Unsplash License

Granite SwitchLLMsOpen ModelsIBM GraniteModel ServingSpeech Recognition