IBM Releases Granite Switch 4.1 With Decision Adapters and Fast Speech Recognition

IBM Granite said on Hugging Face on 15 October 2026 that Granite-Switch-4.1-8B and Granite-Switch-4.1-30B are available as composed models that move between natural-language chat, general and specialised decisions, and speech-to-text from one hosted endpoint.

Granite Switch, introduced earlier in 2026, is IBM’s modular approach: a base Granite model plus Activated LoRA (aLoRA) adapters that share the base KV cache so switching tasks does not mean loading a second full model.

What ships in 4.1

According to the IBM Granite blog, the 8B and 30B Switch 4.1 builds include:

1. A new Decision aLoRA for general-purpose decision-making (e.g. v1/systemone-style tasks)
2. A specialised Guardian aLoRA that, IBM claims, on average outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety-related decision tasks
3. Granite-Speech-5.0-Turbo-CTC, which IBM calls “the fastest ASR model on the OpenASR Leaderboard” — IBM’s claim, tied to that leaderboard at the time of writing

IBM says all three technologies, with the underlying base model, take up approximately the same VRAM as the base model alone, because Switch adapters are cache-efficient Activated LoRAs: the KV cache from a base generation can be reused for subsequent adapter-based decisions. The post also cites recent vLLM work and a co-located ASR encoder for faster switching among chat, decision and spoken requests.

A Hugging Face Space demo is linked for speech-to-decisions on the 8B model: transcribe audio, summarise, screen for harmful content and test user-defined flags on a single endpoint.

Compose-your-own Switch

IBM stresses that 4.1 ships “batteries included” but that anyone can compose a custom Granite Switch by mixing adapters from Granite Libraries. Specialised classification adapters can sit beside a generalist decision adapter at little extra cost because they share the KV cache. The company plans an Adapter SDK for training adapters in Mellea, with more tutorials to follow.

Doom demo footnote — IBM’s speed claim

In a footnote, IBM says a custom 3B Granite Switch model playing Doom made 60–100 decisions per second, compared with Jev’s reported 10 decisions per second, citing a TypeSafe blog post for the Jev figure. That comparison is IBM’s claim about its demo, not an independent bake-off on identical hardware. Unlike Jev in IBM’s telling, the Granite Switch Doom model can also narrate and answer questions while playing.

A separate “Race” anecdote in the post says Granite-Switch-4.1-30B with four specialised aLoRAs completed a five-step RAG workflow faster than Jev-27B with one generic decision LoRA — again, vendor framing.

Why it matters

Decision models — small classifiers that answer structured yes/no or multi-label questions for agent control planes — are becoming a product category alongside chat LLMs. Packing a generalist Decision adapter, a Guardian safety specialist and a fast ASR encoder into one VRAM footprint is IBM’s pitch for voice-in, decide, chat-out agents without a GPU farm per modality.

TSN readers following OpenAI’s Decisions API and Perplexity’s open decider will recognise the same problem: agents need cheap, low-latency structured judgements, not only essays.

What’s still unknown

• Independent OpenASR Leaderboard snapshots confirming the “fastest” ASR claim after publication day
• Pricing and hosting options for Switch 4.1 on IBM’s cloud versus self-host
• Release date for the Adapter SDK in Mellea

What to watch

Watch Hugging Face model cards for Granite-Switch-4.1-8B/30B, the Speech-to-Decisions Space, and IBM’s promised follow-up data dump on Switch benchmarks.

What this means

• One endpoint for chat, decisions and speech. That is the 4.1 product story.
• Leaderboard and outperformance figures are IBM’s. Guardian wins and “fastest ASR” should be read as vendor claims.
• Composable adapters. Libraries + planned SDK matter as much as the prebuilt 8B/30B SKUs.

Sources

1. Hugging Face blog (IBM Granite), Kate Soule & Luis A. Lastras, “Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint”, 15 October 2026
2. TypeSafe, “Introducing System One models and Jev” (cited by IBM for Jev decisions/second figure)

Related stories

• OpenAI’s Decisions API and Perplexity’s Open Decider: AI That Answers With a Probability, Not an Essay
• NVIDIA Releases Nemotron 3.5 Super VL, an Open 120B Multimodal MoE for Agents