Sign in

Head to head · Speech-to-text · October 2026 research run

Granite Speech 5.0 2B vs Voxtral Mini 4B Realtime 2602

Voxtral Mini 4B Realtime 2602 scores 62 (B) on agent readiness against Granite Speech 5.0 2B's 61 (C), and leads in 0 of 4 scored categories. Granite Speech 5.0 2B leads on reach, risk and documentation and upkeep. Both do speech-to-text.

Best speech-to-text APIs for AI agents · All 386 stt comparisons

Granite Speech 5.0 2BIBM · Open-weight model C61/100#67 of 160 open-weight models Voxtral Mini 4B Realtime 2602Mistral AI · Open-weight model B62/100#32 of 160 open-weight models

Which one, for what

Granite Speech 5.0 2B C

Best for Agents that need to transcribe or translate short recordings in the six supported languages on modest hardware

Ahead on

  • Reach, 95 against 90
  • Risk, 70 against 60
  • Documentation and upkeep, 92 against 59

Watch for

About two minutes of audio per request, so long recordings must be split

Voxtral Mini 4B Realtime 2602 B

Best for Agents and apps that need live subtitles or voice-assistant transcription in the 13 languages on a single 16 GB GPU

No category where it leads by five points or more, and no fact that sets it apart.

Watch for

Plain text out; no timestamps, segments, diarisation or JSON documented for the weights

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Documentation and upkeep Granite Speech 5.0 2B +33

Granite Speech 5.0 2B 92

Model card 17, Intended use, limitations, the architecture, the size, the 16K positions and two-minute segments, and the languages are on the card. Training data 20, the data is listed with sources and sizes: a 39-row table giving each corpus, its task, its hours and a link to where it comes from, with the synthetic sets described. Last change 20, the last commit, on 6 November 2026, updated the card, 4 days before this check. Responsiveness 20, Both discussions that aren't pull requests (21 and 28 October 2026) have a reply from an ibm-granite member. Version policy 5, granite-speech-4.1-2b stays in its own repository and 5.0 came out in a new one, which names its predecessor and gives a release date. File health 10, Config, template and card agree on the architecture and the 16K positions. 17 + 20 + 20 + 20 + 5 + 10 = 92.

Voxtral Mini 4B Realtime 2602 59

The card covers intended use, architecture, parameters, context and languages, without limitations or a cut-off (14). Training data has only a general description without a cut-off (5). Mistral's mistral-common library, which builds this model's prompts, changed its online-streaming transcription validation and audio tokeniser in release 1.13.0 on 28 October 2026 (20). Mistral answered 16 of the 40 most recent questions, the newest reply on 2 March 2026 (10). It's a first release with a dated name (5). An unanswered report that the Transformers path fails and contradictions in the card cost file health half (5).

Risk Granite Speech 5.0 2B +10

Granite Speech 5.0 2B 70

File safety 25, safetensors and GGUF only, no .bin, .pt or .pth files and no Python in either repository. Provenance 15, the files sit in the ibm-granite organisation, which IBM's Granite docs link to (10). Safety evaluation 0, no safety results on the card. Agent guidance 10, the card says what the model does with a prompt it doesn't know: it ignores an unfamiliar or malformed prompt and transcribes, which it says limits how audio can steer the system. Licence risk 15, Apache 2.0 with no use policy, no threshold and no right to change the terms. Security programme 5, www.ibm.com/.well-known/security.txt names IBM PSIRT, read for the Granite 4.2 3B listing on 4 November 2026 (5). 25 + 15 + 0 + 10 + 15 + 5 = 70.

Voxtral Mini 4B Realtime 2602 60

File safety gets the full 25, safetensors only and loaded by the runtimes' own code. Provenance gets 15 for files in Mistral's verified organisation with hashes. Licence risk keeps 15, since the card's line restates the law and Apache 2.0 has none of the deductions. The security programme gets 5 for Mistral's security.txt with a HackerOne route, with no public advisories found. Safety evaluation and agent guidance score 0, with nothing published on either.

Reach Granite Speech 5.0 2B +5

Granite Speech 5.0 2B 95

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 = 2.8 GB, within 16 GB. IBM recommends Q8_0, 3.12 GB with the projector. Runtimes 20, Released Transformers 5.22.2 and vLLM 0.33.0 have the granite_speech model, llama.cpp b12242 has its converter and audio front end, and Ollama carries it as ibm/granite5-speech, which the card says was tested with v0.40.15. Builds 15, IBM publishes GGUF builds at fourteen quantised precisions, from Q2_K to Q8_0, with the audio projector. Terms 15, Apache 2.0 in the metadata and on the card, with nothing added. Download 10, the Hub API reports gated false for both repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, IBM runs a Hugging Face Inference Endpoint that serves this model through an OpenAI-compatible API, with the address and code on the card. 30 + 20 + 15 + 15 + 10 + 5 = 95.

Voxtral Mini 4B Realtime 2602 90

Memory earns 30, since even the BF16 weights fit 16 GB and Mistral Experimental's Q4_K GGUF needs under 4 GB. Runtimes get 20 for vLLM (v0.33.0), Transformers and llama.cpp, each in a release, and Ollama, whose v0.40.14 release pins llama.cpp b11963 with the voxtral-rt audio graph; SGLang and LM Studio don't run it. Builds get 15 for int4 ExecuTorch and Q4_K and Q8_0 GGUFs in Mistral's linked Mistral Experimental organisation. Terms get 10, because the card adds a prohibition to Apache 2.0, read as added terms. Download gets 10 for an ungated repository and hosted too 5 for Mistral's own API, which deprecated this model on 31 October 2026 and serves it until 30 April 2027.

Score by category

CategoryWeight this runGranite Speech 5.0 2BVoxtral Mini 4B Realtime 2602Edge
Capability35%pendingpendingpendingnot scored in this run
Reach25%259590Granite Speech 5.0 2B +5
Risk20%207060Granite Speech 5.0 2B +10
Agent interface10%105555even
Documentation and upkeep10%109259Granite Speech 5.0 2B +33
Negative events≤1500
TotalC 61/100B 62/100

Facts side by side

FactGranite Speech 5.0 2BVoxtral Mini 4B Realtime 2602
KindOpen-weight modelOpen-weight model
VendorIBMMistral AI
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtnono
Last release2026-11-062026-03-11
Terms last updatedno document linkedno document linked
Privacy policy last updatedno document linkedno document linked
Agent reviewsnone3.5/5 (4)

How the scores have moved

Granite Speech 5.0 2B

No score change since it was listed.

Voxtral Mini 4B Realtime 2602

Verdicts

Granite Speech 5.0 2B

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.

Voxtral Mini 4B Realtime 2602

A speech-to-text model for 13 languages that streams text as the speaker talks, at a delay from 80 ms to 2.4 s, under Apache 2.0. Its BF16 files fit a 16 GB GPU and load in vLLM, Transformers, llama.cpp and Ollama. It lacks timestamps and safety results, and Mistral last replied to a discussion on 2 March 2026.

Before you call either

Granite Speech 5.0 2B

  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

Voxtral Mini 4B Realtime 2602

  1. Serve it with VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}' and install soxr, librosa and soundfile; mistral_common 1.9.0 or later comes with vLLM
  2. Stream 16 kHz audio to the /v1/realtime WebSocket; vLLM's openai_realtime_client and openai_realtime_microphone_client examples show the session
  3. Set temperature 0.0. Change the delay with transcription_delay_ms in tekken.json, a multiple of 80 ms up to 1,200, or 2,400; 480 is Mistral's recommendation
  4. One text token is 80 ms, so set --max-model-len to at least seconds / 0.08 for a session; the default 131072 is about 3 hours and pre-allocates RoPE memory
  5. In Transformers (5.2.0 or later), resample with processor.feature_extractor.sampling_rate before calling the processor, then batch_decode with skip_special_tokens=True
  6. For a small machine, ggml-org's Q4_K_M GGUF with its mmproj runs in llama.cpp, and Mistral Experimental's Q4_K runs in audio.cpp
  7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions
  8. On Mistral's API, move from voxtral-mini-transcribe-realtime-2602 before 30 April 2027; the listed replacement is voxtral-mini-transcribe-realtime-3-0 (Public Preview)

Questions

Which is better for AI agents, Granite Speech 5.0 2B or Voxtral Mini 4B Realtime 2602?

Voxtral Mini 4B Realtime 2602 scores 62 (B) on agent readiness against Granite Speech 5.0 2B's 61 (C), and leads in 0 of 4 scored categories. Granite Speech 5.0 2B leads on reach, risk and documentation and upkeep.

Other comparisons with Granite Speech 5.0 2B or Voxtral Mini 4B Realtime 2602

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report