Home / Granite Switch

Granite Switch

Modular post-training for Granite 4.1. Each capability for reasoning, safety and agentic workloads is trained on its own as an adapter in the Granite libraries, and Granite Switch composes them into one model that serves chat, decisions and speech from a single endpoint. I lead the Guardian and Core libraries and the general-purpose Decision adapter.

Try the speech demo Granite Switch on GitHub Switch models Release blog, Oct 2026

on IFEval over the best other decision model, from the requirement-check adapter: +17 pts
IFEval balanced accuracy, Granite 4.1 8B prompted vs post-trained: 0.57 → 0.86
decisions from a 3B Switch model playing Doom, against 10 a second reported for Jev: 60–100 / s
decisions on live speech from one endpoint, in the speech demo: 13 in < 0.2 s

Composition

From libraries to one model

Granite libraries

Guardiansafety, policy, factuality, tool calls Corerequirements, uncertainty, attribution RAGanswerability, hallucination, query clarification

Granite Switch 4.1, 8B or 30B

Granite 4.1 base modelDecision adapter, for any yes/no questionLibrary adapters, chosen per requestSpeech encoderAbout the same GPU memory as the base model alone. Adapters reuse the base model’s KV cache.

One endpoint

ChatDecisionsSpeech to text

The 4.1 release ships with a set of adapters built in; anyone can compose their own Switch model from the library adapters.

Results

Against other decision models

Scoreboard comparing the best Granite adapter (blue dot) with four other decision models on 11 benchmarks. Granite leads on 8. — Best Granite adapter (blue) against four other decision models: Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash. Ahead on 8 of 11 benchmarks, by up to 17 points; behind on LLM-AggreFact and InfoBench.Granitelib evaluations, Hugging Face Space ↗

Live

Thirteen decisions from one endpoint

The adapters share the base model’s KV cache, so each extra check costs almost nothing. In the demo, the Decision adapter answers every yes/no question about live speech while the Guardian adapter screens it, together in under 0.2 seconds.

How it works, on the Hugging Face blog

The Speech to Decisions demo: a pizza-order transcript on the left; on the right, a safety check reading Nothing harmful and yes/no questions with scores. — A pizza order, transcribed, labeled and screened as it is spoken.Speech to Decisions, Hugging Face Space ↗