IBM Granite Switch 4.1 8B and 30B fold chat, decisions and speech-to-text into one endpoint
IBM has released Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, models that bundle a Decision adapter, a Guardian safety adapter and a speech-to-text model behind a single endpoint. IBM says the combined model uses roughly the same GPU VRAM as the base model alone. Benchmark claims are IBM's own, and context window and pricing were not disclosed.
IBM has released Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, two models that combine natural-language chat, decision-making and speech transcription in a single hosted model. According to IBM, the composed model needs approximately the same GPU VRAM as the underlying base model by itself. The announcement was published October 15, 2026 on the Hugging Face blog.
What's in the bundle
Both models are "Granite Switch" builds: a frozen Granite base model compiled together with a predefined set of adapters. IBM lists three added components:
- Decision aLoRA: a new Activated LoRA for general-purpose decision tasks.
- Guardian aLoRA: a specialized safety adapter. IBM claims it outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on average across safety-related decision tasks. No scores were included in the post, and the comparison has not been independently verified.
- Granite-Speech-5.0-Turbo-CTC: an embedded ASR model that IBM describes as the fastest on the OpenASR Leaderboard.
IBM says the 8B model can transcribe audio, generate summaries, screen for harmful content and test 15+ generic flags on one endpoint in real time.
How it works
The adapters are built as Activated LoRAs (aLoRAs), which are designed to be cache-efficient. The KV cache produced by a base-model generation can be reused by any later adapter-based decision, so a chat turn followed by a classification does not require re-processing the prompt. IBM says this, together with its recent vLLM contributions that speed up low-rank adapter inference and a co-located ASR encoder, lets the models switch between chat, decision and spoken requests quickly.
Because adapters share the base model's KV cache, IBM argues that adding specialist adapters alongside the generalist Decision adapter "costs almost nothing." The specialists come from the earlier Granite Libraries release, and IBM claims many of them beat generalist decision models on specialized classification tasks.
Comparison with Jev
IBM benchmarked Granite-Switch-4.1-30B (frozen Granite base plus four specialized aLoRAs) against Jev-27B (frozen Qwen base plus one generic decision LoRA) on a real-world, 5-step RAG workflow. IBM says the Granite model finished much faster, but it did not publish timings in the post.
In a separate demo, IBM built a custom 3B-parameter Granite Switch model to play Doom. It reports 60–100 decisions per second, against 10 decisions per second that Jev's developers reported. IBM says the model can also narrate gameplay and answer questions while playing. This 3B model is a demonstration and is not part of the 4.1 release.
Build your own
Users do not have to use IBM's prebuilt versions. Anyone can compose a Granite Switch model from adapters in Granite Libraries. IBM plans to release an Adapter SDK for training adapters in Mellea, plus tutorials and further Granite Switch support. More data and details are promised "in the upcoming weeks."
Not disclosed
- Context window: not disclosed in the announcement
- Pricing: not yet disclosed; the post covers model releases rather than a priced API
- Training cutoff and licensing terms: not stated in the post
- Quantitative benchmark scores: none published in the post
What this means
The pitch is operational rather than about raw model quality. Teams running separate chat, guardrail, classifier and ASR services pay for each in GPU memory and in repeated prompt processing. Sharing one base model and one KV cache across adapters targets both costs. If the VRAM claim holds under production load, it matters most for agentic and RAG pipelines, where every step ends in a small classification.
The evidence so far is thin. The Guardian comparison and the Jev race are IBM's own, with no published numbers, and the competing systems differ in base model and adapter count. The Doom comparison also sets a reported figure from another vendor against IBM's own run. Wait for the promised data and independent testing before treating the accuracy or latency claims as established.
Related Articles
Meta releases Muse Spark Open, an 800B-parameter open-weight model built for Apple M5 Ultra Macs
Meta released Muse Spark Open, an 800-billion-parameter open-weight model, on October 15, 2026. It is optimized for Apple's M5 Ultra Mac Studio and needs roughly 1TB of memory to run locally. Meta claims it outperforms GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.
Nvidia releases Nemotron 3.5 Super VL: open 120B multimodal MoE with 12B active parameters
Nvidia has released Nemotron 3.5 Super VL, an open multimodal Mixture-of-Experts model with 120 billion total and 12 billion active parameters, built for continuously running enterprise agents. Two compact Bolt models for local PCs and edge devices launch alongside it, though the 4 GB variant ships later this month.
Mistral's 209M-parameter Timestral tops GIFT-Eval, TIME and fev-bench time-series leaderboards
Mistral has released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. Mistral claims the largest model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, and runs on CPUs without a GPU.
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open-weights image model
FLUX 3 Image Klein is a 5.6B-parameter open-weights image model covering text-to-image, multi-reference generation and image editing. Its developer claims it leads its own blind evaluations in all three tasks. Pricing, license terms and benchmark data were not disclosed in the announcement.
Comments
Loading...