IBM Releases Granite Switch 4.1: Chat, Decisions and Speech on One Endpoint

IBM shipped public 8B and 30B checkpoints

IBM’s repository cards list October 14, 2026 as the release date for Granite Switch 4.1, and IBM authors published the launch article on Hugging Face on October 15. At review time, ibm-granite/granite-switch-4.1-8b and ibm-granite/granite-switch-4.1-30b were public and ungated. The API records exposed BF16 parameter metadata of 8,776,400,896 and 29,989,814,272 parameters respectively.

These are inspectable model artifacts, not a promise of weights later. Both repositories provide sharded safetensors, configuration, tokenizer files, chat-template logic, adapter mappings and composition records. The cards specify Apache 2.0 licensing, a 131,072-token context inherited from the base, and Transformers and vLLM as supported backends. They do not publish a universal managed API price, quota or service level.

One checkpoint contains four different operating modes

The base model handles ordinary text generation. A decision aLoRA accepts a state plus labelled options and returns a probability distribution over option labels. A guardian-core aLoRA checks selected safety categories and emits a yes-or-no JSON result. An embedded 470M-parameter Granite Speech 5.0 TurboCTC encoder transcribes English audio, after which the transcript is inserted into the text context.

Calling this one model should not erase the component boundaries. The adapter is selected through a control token inserted by the chat template, and IBM describes the design as coarse-grained switching: one adapter is active across decoder layers for a contiguous token span. The speech path is a cascade. It does not make the Granite language model natively reason over acoustic features.

The speech path is convenient and deliberately limited

IBM describes Granite Speech 5.0 TurboCTC as a 470M-parameter, English-only conformer CTC encoder trained on about 60,000 hours of public or public-corpus-derived audio. It uses non-autoregressive greedy decoding. The Switch configuration enables audio, accepts clips through an OpenAI-compatible chat request and places the resulting transcript where the audio content appeared.

The model card is unusually direct about the boundary: the speech encoder and LLM are not jointly trained, there is no learned projection coupling them, output is lowercase and unpunctuated, and there is no decode-time steering. A transcription error can silently alter a later summary, routing choice or safety verdict. IBM’s “fastest ASR” and leaderboard statements are dated vendor reports, not measurements made for this article.

Activated LoRA is the infrastructure bet

IBM says activated LoRA lets a request reuse the base model’s key-value cache when it later switches into a decision adapter. Both adapters are embedded in the checkpoint, and the card says the combined technologies occupy approximately the same GPU memory as the base model. That could reduce the number of separately loaded models and servers in a pipeline.

Those are architecture and provider-performance claims, not a sizing guarantee. Teams should measure checkpoint load time, usable context, prefix-cache hit rate, adapter-switch overhead, concurrency, tail latency and peak memory on their own runtime. The 8B card requires PyTorch 2.11 or newer, Transformers 5.16 through 5.17, and optional vLLM 0.26 through 0.27; version pinning is part of the deployment decision.

Published decision scores need workload validation

IBM reports evaluating the decision adapter on 1,068 rows from the public Decision Index fast subset under edition-v0.3 scoring. The cards report overall scores of 42.9 for the 8B checkpoint and 50.1 for the 30B checkpoint, with locally served and hosted-vLLM paths said to agree within about 0.6 points. These are vendor-run results using a specific prompt and label-logit recipe.

The adapter returns probabilities over supplied labels, not a certified business decision. Calibration can move across domains, class definitions and base rates. For a real routing or approval workflow, preserve an abstain path, compare confidence with observed accuracy, test rare and adversarial cases, and keep consequential actions behind policy and human review. The guardian adapter is likewise an additional signal, not an application safety system.

What is available now—and what is not

Available now are public Apache 2.0 checkpoints, composition code, an OpenAI-compatible vLLM serving example, Transformers support and a Hugging Face demonstration. Builders can reproduce the composed model or serve either checkpoint themselves. The repositories do not establish production capacity, a managed IBM endpoint, regional availability, data-processing terms, uptime or token pricing.

Also absent from the current checkpoints are the broader Guardian Library’s factuality detection, factuality correction and policy-guardrail adapters. IBM says the decision adapter has no Mellea intrinsic yet and calls its Adapter SDK planned work. Buyers should separate today’s downloadable artifacts from those future integrations.

What builders should do next

Trial the 8B checkpoint first when consolidation is the goal. Pin its commit, run one fixed runtime stack, and test four paths separately: ordinary chat, decision labels, guardian checks and noisy speech. Then test the handoffs, especially whether transcription errors or stale cached context change decisions. Record output quality, calibration, memory, latency, cache reuse and total cost per completed workflow.

Move to 30B only if paired tasks justify the extra serving footprint. Constrain the system when labels are ambiguous, audio is noisy or the Guardian result would trigger a consequential action. Wait if a managed-service contract or broader adapter set is required. Reject the design if one shared endpoint creates a failure domain that outweighs the operational savings.

Disclosures

AccessAllGPT received no weights, API access, credits, hardware, data, benchmark logs, private briefing, review, payment or compensation for this article. IBM and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Granite Switch 4.1. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with IBM or Hugging Face. Publication-wide relationships are listed on the disclosures page.