Apache-2.0llmRuns locally8.78B (8B card safetensors)
Granite Switch 4.1: Specs, License & How to Run
IBM Apache-2.0 modular 8B+30B Switch — chat, decision aLoRA, embedded ASR.
- License
- Apache-2.0
- Parameters
- 8.78B (8B card safetensors)
- HF likes
- 55 (2026-11-01)
- First seen
- 2026-10-14
TL;DR. Granite Switch 4.1 is a llm model from IBM. Parameters: 8.78B (8B card safetensors).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs, LLM classification APIs.
| Developer | IBM |
|---|---|
| Type | llm |
| Architecture | granite_switch / GraniteSwitchForCausalLM. 8B card: hidden 4096, 42 layers, vocab 100355, max_position_embeddings 131072; adapter_names guardian-core (rank 16) and decision (rank 64). 30B card: hidden 4096, 66 layers, same vocab and 131072 context. Base models ibm-granite/granite-4.1-8b and granite-4.1-30b plus aLoRA adapters and embedded granite-speech-5.0-470m-turboctc. Distinct from the dedicated Granite Speech 5.0 page. |
| Parameters | 8.78B (8B card safetensors) |
| Modalities | text, audio, speech |
| Languages | en, fr, de, es, pt, it, ja |
| Context | 128K (card); 131,072 tokens (config) |
| License | Apache-2.0 — Hugging Face cardData.license is apache-2.0 on granite-switch-4.1-8b and granite-switch-4.1-30b (2026-11-01). The 8B card has no LICENSE blob (404). The compose repo generative-computing/granite-switch is Apache-2.0 on the GitHub license API and LICENSE file. Not legal advice. |
| Hardware | Not yet verified — 8B usedStorage 35,108,465,504 bytes (~32.7 GiB); safetensors.parameters 8,776,400,896. 30B usedStorage 119,962,243,488 bytes (~111.7 GiB); safetensors.parameters 29,989,814,272. Peak VRAM is not independently measured. Card backends: Transformers and vLLM via granite-switch>=0.1.1. (not independently tested) |
| Hugging Face likes | 55 (as of 2026-11-01) |
| Model card | https://huggingface.co/ibm-granite/granite-switch-4.1-8b |
| First seen | 2026-10-14 |
Granite Switch 4.1 is IBM’s modular Granite 4.1 compose: one hosted endpoint that moves between natural-language chat, a decision aLoRA, Guardian Core, and embedded speech-to-text. Official hosts we opened on 2026-11-01: ibm-granite/granite-switch-4.1-8b (created 2026-10-14T21:38:09Z; last modified 2026-10-28T20:44:15Z; 55 likes / 4,204 downloads), ibm-granite/granite-switch-4.1-30b (created 2026-10-14T21:23:12Z; last modified 2026-10-28T20:45:20Z; 44 likes / 3,185 downloads), the 4.1.1 blog Granite Switch 4.1.1, the 4.1 blog Granite Switch 4.1, and github.com/generative-computing/granite-switch (119 stars / 20 forks; Apache-2.0). API gated was false on both cards.
This identification page hangs off ChatGPT alternatives, closed LLM APIs, and classification / decision APIs for the decision adapter. Those mappings are editorial. We have no sourced bake-off against ChatGPT or Clef.
Dedicated speech family (different URL): Granite Speech 5.0 on Otter.ai alternatives — Nemotron 3.5 Transcribe vs Granite Speech 5.0. Switch embeds granite-speech-5.0-470m-turboctc; it does not replace that page.
Family split (do not mix the ids)
| Card | Base (card) | safetensors total | Likes / downloads (2026-11-01) |
|---|---|---|---|
| granite-switch-4.1-8b | granite-4.1-8b, 128K | 8,776,400,896 (~8.78B) | 55 / 4,204 |
| granite-switch-4.1-30b | granite-4.1-30b, 128K | 29,989,814,272 (~30.0B) | 44 / 3,185 |
Both main branches are revision 4.1.1 (card release October 27, 2026). Original 4.1 is October 14, 2026. Same two adapters: guardian-core and decision. 4.1.1 speech locales on the cards: English, French, German, Spanish, Portuguese, Italian, Japanese.
Preview siblings named on the cards (no audio; not this page): granite-switch-4.1-8b-preview, -30b-preview, -3b-preview.
What we verified (2026-11-01)
| Field | 8B | Source |
|---|---|---|
| Host | huggingface.co/ibm-granite/granite-switch-4.1-8b (not gated) | HF API |
| Created / last modified | 2026-10-14T21:38:09Z / 2026-10-28T20:44:15Z | same |
| License tag | apache-2.0 | HF cardData |
| LICENSE blob | 404 | .../resolve/main/LICENSE |
| Pipeline tag | text-generation | HF API |
| Architecture | GraniteSwitchForCausalLM |
HF config |
| usedStorage | 35,108,465,504 bytes | HF API |
| Compose repo stars | 119 | GitHub API |
Card (copied): speech is not jointly trained with the LLM — there is no trained projection coupling them yet; a mistranscription can change what chat or an adapter sees. Guardian Core on this checkpoint does not include the Guardian Library’s factuality-detection, factuality-correction, or policy-guardrails adapters.
How to run (from the official README, not tested)
pip install "granite-switch[vllm,audio]>=0.1.1"Cards also name granite-switch[hf] and granite-switch[compose]. vLLM serve uses the Hugging Face id. forkalt has not served either checkpoint. See vLLM.
Vendor benches (self-reported)
JEV Decision Index v0.3 overall (card, ×100, chance-corrected weighted aggregate): 46.72 (8B) and 51.33 (30B). Those rows stay self-reported. They are not a Cloudflare Clef or Mammoth-1 ranking — Mammoth-1 vs Cloudflare Clef.
License
Card tag apache-2.0. No LICENSE file on the 8B siblings list we opened. granite-switch LICENSE is Apache-2.0. Not legal advice.
Related: Granite Speech 5.0 — Nemotron 3.5 Bolt — LFM3-1.7B — LFM3 — Clef / Mammoth-1 on classification APIs. Catalog: home. Corrections: contact.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| JEV Decision Index v0.3 overall (8B decision aLoRA, card) | 46.72 (self-reported) | https://huggingface.co/ibm-granite/granite-switch-4.1-8b |
| JEV Decision Index v0.3 overall (30B decision aLoRA, card) | 51.33 (self-reported) | https://huggingface.co/ibm-granite/granite-switch-4.1-30b |
Deploy options
FAQ
What is Granite Switch 4.1 versus revision 4.1.1?
Both 8B and 30B cards say main is revision 4.1.1 (release date October 27, 2026). Revision 4.1 (October 14, 2026) is a tagged pin. 4.1.1 keeps the same two adapters and updates decision weights plus multilingual TurboCTC speech (7 languages). Load the older tag with revision=granite-switch-4.1-8b (or -30b).
Are the weights Apache-2.0?
Card tags apache-2.0 on both family cards (2026-11-01). The 8B LICENSE blob 404s. github.com/generative-computing/granite-switch is Apache-2.0 (119 stars). Not legal advice.
Is this the same as Granite Speech 5.0?
No. Granite Speech 5.0 is the dedicated ASR/AST family. Switch embeds granite-speech-5.0-470m-turboctc so one endpoint can transcribe, then chat or decide. The speech encoder and the LLM are not jointly trained (card). See Granite Speech 5.0.
Does it replace ChatGPT or a classification API?
No sourced bake-off. Chat, decision aLoRA, and Guardian Core share one composed checkpoint. JEV Decision Index rows stay self-reported.
Sources & verification
- https://huggingface.co/ibm-granite/granite-switch-4.1-8b
- https://huggingface.co/ibm-granite/granite-switch-4.1-30b
- https://huggingface.co/ibm-granite/granite-switch-4.1-8b/raw/main/config.json
- https://huggingface.co/blog/ibm-granite/granite-switch-4-1-1
- https://huggingface.co/blog/ibm-granite/granite-switch-4-1
- https://github.com/generative-computing/granite-switch
- https://github.com/generative-computing/granite-switch/blob/main/LICENSE
- https://research.ibm.com/blog/granite-libraries-project-switch
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.