IBM announced Granite Switch 4.1 on October 15, 2026, bringing chat, structured decision scoring, safety checks and English speech transcription into a shared model-serving package. The 8B and 30B model cards list October 14 as their release date and specify a 128K context window and Apache 2.0 licensing. 1 2 3
The central development is deployment consolidation—not a jointly trained speech-language model. The project’s audio documentation describes an alpha transcription cascade: a speech recognizer converts recordings into text, which the language model and specialized adapters then process. That distinction matters when assessing both the capabilities and IBM’s performance claims. 4
What the package brings together
According to IBM’s model cards, both configurations combine a corresponding Granite 4.1 language model with two adapters, decision and guardian-core, and the same English speech recognizer. The earlier preview configurations had a broader adapter bundle but no audio, so their capability lists should not be treated as interchangeable with this release. 2 3
The components have distinct roles:
- Chat and text generation: The base language model handles ordinary conversation. The Decision adapter is not itself a chat model. 5
- Structured decisions: IBM’s Decision documentation describes scoring caller-supplied options using next-token logits, supporting choice, yes/no and ordinal-score questions. Each question requires an independent pass over the supplied state. 5
- Safety and groundedness checks: Guardian Core evaluates safety risks, retrieval-augmented generation groundedness and function-call hallucination criteria through prescribed scoring templates, rather than unrestricted generation. 6
- Speech input: Granite Speech 5.0 TurboCTC is documented as a 470-million-parameter English automatic speech recognition model. Its role is transcription, not speech generation. 7
The default checkpoint does not contain the entire Guardian Library. Separate factuality-detection, factuality-correction and policy-guardrails adapters are absent, although Guardian Core retains its own groundedness-scoring capability. 3 6
IBM’s one-endpoint framing is best understood as a shared serving deployment, not proof that every capability uses an identical HTTP route. The model card demonstrates audio through chat completions, while the announcement separately discusses /v1/systemone compatibility. 2 1
How switching works—and where speech stops
The specialization mechanism is activated LoRA, or aLoRA. The underlying paper describes adapters that change processing only after an invocation point. Earlier tokens retain the base model’s key-value cache, avoiding the full-prefix recomputation associated with switching to a conventional LoRA adapter. That supports reuse of prior computation; it does not make switching computation-free or arbitrary caches interchangeable. 8
This architecture predates the October package. IBM introduced Granite Libraries and Project Granite Switch on June 4, 2026, and the speech model’s card lists an August 25 release date. Switch 4.1 therefore integrates previously introduced components rather than debuting the modular concept itself. 9 7
For audio, the processor replaces the audio marker with transcript tokens before passing them to the language model. The decoder and adapters operate on text, not directly on learned audio embeddings. The default recognizer produces English-only, lowercase, unpunctuated transcription, and the documentation identifies no decode-time steering for it. 4
That creates an important downstream risk: IBM’s release card warns that mistranscription can silently change what the language model or adapter evaluates. A safety check on an incorrect transcript is not necessarily a safety check on what the speaker actually said. 3
What IBM’s benchmarks establish
IBM’s release cards report these results on the Decision Index v0.3 public fast subset, comprising 1,068 rows, under hosted serving: 3 2
Configuration IBM-reported overall score Granite Switch 4.1 8B 42.9 Granite Switch 4.1 30B 50.1These are public-component scores on a 0–100 scale—not general accuracy percentages or demonstrated full-leaderboard rankings. 3 2
A separate Decision adapter README reports 42.56 for 8B and 50.91 for 30B, describing its evaluation as the v0.3 public set. The release cards specify a fast subset and hosted-serving setup. The reviewed documentation does not fully reconcile those differences, so the two sets of totals should not be combined or treated as interchangeable. 5 3 2
For safety, IBM reports average F1 scores of 0.80 for Guardian 4.1 8B aLoRA and 0.81 for the 30B version across ten out-of-distribution benchmarks. Its comparative table does not show a win on every dataset and provides no confidence intervals. The small aggregate margins therefore do not establish statistically significant superiority. These remain IBM’s measurements, not independent evaluations. 6
The workflow-speed evidence is less complete. IBM says a 30B Switch configuration with four specialized aLoRAs completed a five-step retrieval-augmented generation workflow faster than Jev-27B with one generic decision LoRA. The announcement supplies neither numerical elapsed times nor hardware details. It also tests a different adapter configuration from the shipping two-adapter package, preventing a reproducible speedup claim for the default release. 1
IBM also characterizes the speech recognizer as the fastest ASR model. Its model card ties the Open ASR results to an October 2, 2026 snapshot, with throughput measured on one H200 using short-form English benchmarks—not complete speech-to-decision latency. The supplied research retrieved the leaderboard landing page but not its underlying result rows, so it did not independently confirm a current first-place ranking. 1 7 10
Deployment and cost remain workload questions
IBM describes Switch as using approximately the base model’s VRAM. The 8B build manifest records 8,380,551,168 base parameters and 8,776,400,896 composed parameters, a 4.72% increase. That is parameter accounting, not a total-memory benchmark; the audio documentation separately warns that ASR weights require GPU headroom. These records do not establish quantified operating-cost savings. 1 11 4
The retrieved project README documents Hugging Face and vLLM deployment, specifying Python 3.11+, PyTorch 2.11+ and vLLM 0.26.x–0.30.x, with CUDA 13+ for those vLLM versions. Because repository documentation and model cards are mutable, compatibility assessments need to be tied to specific revisions. 12
IBM also claims native llama.cpp support. A related upstream implementation pull request was included in the research, but the review did not establish which released versions support each capability or whether downstream applications offer audio-feature parity. 1 13
What still needs testing
The clearest potential benefit is fewer separately managed services for transcription, generation and scoring. That is an architectural opportunity, not a demonstrated economic outcome: savings depend on workload, concurrency, cache reuse and GPU allocation. 4 9
Several production questions remain open:
- Reproducible speed: What hardware, precision, prompts, cache conditions, raw timings and exact model revisions produced IBM’s RAG comparison? 1
- Concurrent performance: What are peak VRAM, throughput and p50/p95 latency when speech, chat and decision requests run together? 11 4
- Decision robustness: How well calibrated are option probabilities in unfamiliar domains, and how do accents, noise and transcription errors affect downstream safety judgments? 3 4
IBM advises treating decision probabilities as a prior on unfamiliar domains and validating application-specific behavior before deployment. Restricting responses to supplied labels can make an interface more predictable, but a valid label can still be wrong. 3 5
Research scope: This draft uses only the supplied research package, with a cutoff of October 17, 2026. The underlying review examined documentation without executing models or reproducing benchmarks. No independent end-to-end replication was retrieved; multiple IBM-authored artifacts do not constitute independent verification.