# Granite Speech 5.0 2B (slim)

> Granite Speech 5.0 2B is IBM's open-weight speech model of 2.3B parameters for transcription in six languages and speech translation to and from English, with keyword lists. Apache 2.0, with IBM's own GGUF builds from about 2 GB.

- Full: https://www.anchorterminal.com/tools/granite-speech-5-0-2b.md (~6,900 tokens) · this version ~1,830 tokens · JSON https://www.anchorterminal.com/tools/granite-speech-5-0-2b.json · canonical https://www.anchorterminal.com/tools/granite-speech-5-0-2b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/granite-speech-5-0-2b#pricing · reviews https://www.anchorterminal.com/tools/granite-speech-5-0-2b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #67 of 160 open-weight models · #19 in Speech-to-text · not agent-ready · confidence low**


Assessment: A small, well-kept speech model with a clean licence: six languages, translation, keyword lists, IBM's own GGUF and an Ollama entry, a signed repository and training data listed corpus by corpus. It handles about two minutes of audio at a time and returns plain text, with no timestamps or speakers, and publishes no safety results.

## Facts

- Kind: Open-weight model · vendor: IBM · category: Speech-to-text · legal entity: International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian) · provenance 88/100
- Packages: huggingface `ibm-granite/granite-speech-5.0-2b`, huggingface `ibm-granite/granite-speech-5.0-2b-GGUF`, ollama `ibm/granite5-speech`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/ibm-granite/granite-speech-5.0-2b and the GGUF repository beside it, read from the model card, the files and their metadata, the config, the chat template, the commit history, Ollama's page for ibm/granite5-speech and the runtimes' own code at their latest releases on 10 November 2026. IBM's company facts come from our Granite 4.2 listings. The blog post, the papers and the ONNX repository weren't read. Nothing was downloaded or run
- Model: 2,313,207,132 parameters in the files. A 16-block conformer speech encoder trained with CTC, a two-layer window Q-Former that brings the audio down to 10 embeddings a second, and an intermediate checkpoint of granite-4.0-1b-base as the language model, extended to 16K positions. GraniteSpeechForConditionalGeneration
- Languages: English, French, German, Spanish, Portuguese and Japanese for transcription, and translation to and from English for the same, plus English to Italian and English to Mandarin
- Files: Two safetensors shards in BF16 (5.04 GB), chat_template.jinja, config, processor and tokenizer files, a sample recording and model.sig, a sigstore signature. No pickle files and no Python. The GGUF repository has fifteen precisions from Q2_K to BF16 and a 1.16 GB audio projector. Hugging Face lists a SHA-256 for the large files
- Licence: Apache 2.0, in the metadata of both repositories and on the card
- Runtimes: The card documents Transformers (4.52.1 or later), vLLM, llama.cpp with IBM's GGUF, Ollama (ibm/granite5-speech, tested with v0.40.15), mlx-audio 0.5.9 on Apple Silicon and a LiteRT-LM conversion from the LiteRT community. Released Transformers 5.22.2, vLLM 0.33.0 and llama.cpp b12242 carry the architecture. SGLang 0.5.23 doesn't
- Input and output: 16 kHz mono audio of up to about two minutes with a task prompt, and text out. The card's prompt table covers raw or punctuated transcripts, a named language, translation, and a keyword list to bias towards, up to 2,000 terms. No timestamps or speaker labels are documented
- Hardware: No memory figure on the card. Q8_0 with the projector is 3.12 GB and BF16 5.04 GB. The LiteRT-LM conversion is a 2.8 GB file tested on Android and macOS
- Training data: 390,000 hours of audio from public corpora and synthetic sets, per the card, listed in a 39-row table with the task, the hours and a source link for each: Common Voice 17, Multilingual LibriSpeech, LibriHeavy, VoxPopuli, Granary and others. No cut-off date

- Scores: Capability pending, Reach 95, Risk 70, Agent interface 55, Documentation and upkeep 92 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, the smallest build listed that keeps 4 bits is IBM's Q4_K_M GGUF, 1.14 GB with the 1.16 GB audio projector, and 2.30 GB × 1.2 = 2… · Risk, File safety 25, safetensors and GGUF only, no .bin, .pt or .pth files and no Python in either repository. · Agent interface, Chat template 20, chat_template.jinja in the repository (193 characters). · Documentation and upkeep, Model card 17, Intended use, limitations, the architecture, the size, the 16K positions and two-minute segments, and the languages are on th…
- Sources: 14, open questions: 6, both in the full twin
- Capabilities: speech.stt, speech.batch, speech.languages, speech.translation, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/granite-speech-5-0-2b.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://huggingface.co/ibm-granite/granite-speech-5.0-2b/discussions; https://github.com/ibm-granite/granite-4.2-language-models/issues; more in the full twin

## Before you call it

1. Split recordings into segments of up to about two minutes of 16 kHz mono audio before sending them
2. Use the card's prompt for the task word for word, such as `transcribe the speech with proper punctuation and capitalization.`
3. Use Q8_0. IBM says lower precisions lose accuracy on audio over about a minute
4. To spell names and jargon right, add a keyword list to the prompt, up to 2,000 terms
5. With vLLM, serve with `--max-model-len 2048`, as the card does
6. An unknown prompt falls back to transcription, so check that the output is the task you asked for

## Connect

```bash
brew install llama.cpp
llama-cli -st -hf ibm-granite/granite-speech-5.0-2b-GGUF:Q8_0 --audio audio.wav -p "transcribe the speech with proper punctuation and capitalization."
# or: ollama run ibm/granite5-speech:2b, or vllm serve ibm-granite/granite-speech-5.0-2b --max-model-len 2048
```

```bash
# vLLM and llama-server both serve an OpenAI-compatible chat API. Send 16 kHz mono audio of up to about
# two minutes as base64 in the message, with one of the card's task prompts as the text
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Nemotron 3.5 Transcribe 30B-A4B | C | 61 | speech.stt, speech.batch, speech.languages, speech.translation, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-transcribe-30b-a4b.min.md |
| Azure AI Speech speech-to-text | BB | 75.1 | speech.stt, speech.batch, speech.languages, speech.translation | https://www.anchorterminal.com/tools/azure-speech-to-text.min.md |
| Gladia Speech-to-Text API + MCP | B | 69.8 | speech.stt, speech.batch, speech.languages, speech.translation | https://www.anchorterminal.com/tools/gladia-stt.min.md |
| Soniox Speech-to-Text | B | 68.5 | speech.stt, speech.batch, speech.languages, speech.translation | https://www.anchorterminal.com/tools/soniox-stt.min.md |
| Speechmatics Speech-to-Text | B | 67.8 | speech.stt, speech.batch, speech.languages, speech.translation | https://www.anchorterminal.com/tools/speechmatics-stt.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)