HyperWhisper does not lock you into one engine. It ships a library of models, because no single model wins on everything. Each model makes a different trade between privacy, language coverage, speed, accuracy, and cost. This page lists every model and explains why each one is there. There are two kinds of model:

• Speech-to-text — changes your voice into text. This is the transcription step.
• Post-processing — an optional LLM step. The LLM removes filler words, adds punctuation, and formats the transcript after transcription.

You choose both models in the app under Model Library. You can also choose them for each mode in the mode editor.

​

Three ways to run a model

Every model — speech-to-text or post-processing — is in one of three groups:

​

Speech-to-text models

​

On-device

These models run on your machine and make no network calls. After you download a model (where applicable), it works offline. The app never uploads your audio. For more information, see Data Privacy.

​

Whisper models

The general-purpose multilingual models from OpenAI. The VRAM values give the recommended GPU memory for full acceleration. With less memory, the model still runs on the CPU or on part of the GPU, but more slowly.

​

NVIDIA Parakeet models

NVIDIA Parakeet models are faster than Whisper models of an equivalent size. They are accurate for the languages that they support. The Parakeet V3 language list is different on each platform. Each app ships its own export of the model, and the two exports do not carry the same languages.

​

NVIDIA Nemotron 3.5 models

Nemotron 3.5 ASR from NVIDIA is the newest on-device option. macOS gets a broader variant, and Windows gets a smaller streaming variant. On macOS, Nemotron 3.5 is more accurate than the other local models, and it goes far beyond European languages. The macOS multilingual variant is the only local model that supports Chinese, Japanese, Korean, and Arabic.

​

Apple Speech & Qwen3 ASR

• Apple Speech is built into macOS. There is no download, and it is available when you start the app. It is the fastest private option for daily Mac dictation. (It requires a recent macOS version.)
• Qwen3 ASR is an additional multilingual on-device model for users who want the ASR model from Alibaba. It runs on macOS and Windows, and it uses the sherpa-onnx daemon on Windows.

​

Offline language coverage

This table gives the best local model for the most common languages. For the full Parakeet V3 and Nemotron language lists, see the sections above.

​

HyperWhisper Cloud

HyperWhisper Cloud is built into the app. You need no API key and no separate account. It routes to top providers across three accuracy tiers, and you pay only for actual speech (silence and empty recordings cost 0 credits). HyperWhisper Cloud is the best choice when you want the highest accuracy with no setup. HyperWhisper Cloud gives you all 12 supported provider engines with no key. The other providers are Soniox, OpenAI, AssemblyAI, Mistral Voxtral, Gemini, and Meta Muse Voice Transcribe. Only ElevenLabs, Grok, Deepgram, and Groq have automatic failover to a backup provider. The other providers have no automatic failover. For pricing, cost examples, and language guidance, see Providers.

​

Microsoft MAI-Transcribe 2

The Microsoft row of the Model picker gives two models. MAI-Transcribe 2 is a public preview model and the model that the row selects first. It costs approximately 1.67 credits per audio minute and it lists 60 languages. MAI-Transcribe 1.5 stays available at approximately 6.0 credits per minute with 42 languages. Both models are available through HyperWhisper Cloud only, and both accept custom vocabulary as a phrase list. Both models accept WAV, MP3, and FLAC files only, to a maximum of 300 MB. Read File Transcription for the automatic conversion that macOS applies. Microsoft documents speaker diarization and word-level timestamps for MAI-Transcribe 2, but HyperWhisper does not request them. Your transcript thus contains the text only.

​

Meta Muse Voice Transcribe 1.0

Muse is available through HyperWhisper Cloud credits or with your own Meta Model API key. The Cloud route costs 3 credits per audio minute. The Model Library marks code switching, endpointing, context bias, language bias, turn-level timestamps, and speaker diarization as supported. It marks word-level timestamps as not supported. This capability record describes the model. HyperWhisper’s current batch path returns transcript text and does not show all turn metadata. Meta offers real-time streaming upstream, but HyperWhisper does not expose Muse in HyperWhisper Live. Read Providers for the complete record and the official Meta references.

​

Bring your own key

You can use your own key if you hold API credits, if you want the free tier of a provider (Deepgram $200, AssemblyAI $50), or if you need a specific model. Add the key under API Keys. You pay the provider directly at the published rate. Supported providers for bring-your-own-key transcription: OpenAI · Groq · Deepgram · AssemblyAI · ElevenLabs · Mistral · Soniox · Google Gemini · Google Gemini 3.5 Transcribe · SpaceXAI (Grok)

​

Retired speech models

A provider can retire a speech model. HyperWhisper maps the model to a current model of the same provider. It usually removes the retired model from the Model Library and from the Mode editor. A retired row can stay for one release while the migration reaches saved modes. Your saved modes, your per-mode model settings, and your imported backups continue to work. The app applies the map when it reads the mode, thus you do not edit the mode yourself. Both platforms apply the map. The table gives the platform when only one platform has the model. The three retired OpenAI rows remain in the picker for one release. They are not promoted. New OpenAI modes use GPT Transcribe by default, and saved modes move to the replacements in the table. The Deepgram row covers the domain-tuned Nova-2 models (Meeting, Phone Call, Voicemail, Finance, Conversational AI, Automotive, Video) and every Enhanced, Base, and Whisper model. Nova 2 General and Nova 2 Medical are not retired. The Google Chirp 3 row is different from the other rows. Chirp 3 was the Google engine of HyperWhisper Cloud, and Google Gemini 3.5 Transcribe replaced it. The app changes each saved mode one time, at the first start after the update. A mode that used Chirp 3 then uses Gemini 3.5 Transcribe, and it bills the rate of Gemini 3.5 Transcribe. Your imported backups get the same change.

​

Post-processing models

Post-processing is an optional second step. After transcription, an LLM removes filler words and corrects the punctuation and the capitalization. It also applies the formatting that your mode specifies. Post-processing is separate from the speech model, and you can combine any speech model with any post-processing model.

​

Cloud post-processing

You can use cloud post-processing through HyperWhisper Cloud with no key, or with your own API key. Every cloud post-processing model has a speed rating and an accuracy rating. The Rating scale section explains these ratings. They let you choose the trade that is important to you.

​

Gemini 3.8 Flash

Gemini 3.8 Flash is available for post-processing with your own Google Gemini API key and through HyperWhisper Cloud. The model appears in the Gemini model picker on macOS and Windows. On Linux, select it from the HyperWhisper Cloud model list, or enter gemini-3.8-flash for Google Gemini with your own key. It is not the default Gemini model, so select it explicitly when you want to use it. The app rates Gemini 3.8 Flash at 5 for accuracy and 2 for speed. HyperWhisper Cloud disables reasoning for this model, because post-processing needs text cleanup rather than a reasoning response. The Cloud route charges for the input and output tokens. Its introductory rates are $0.75 per 1 million input tokens and$3.75 per 1 million output tokens through December 31, 2026. Gemini 3.8 Flash is a post-processing model. It is different from Gemini 3.5 Transcribe, which changes audio into text.

​

Retired post-processing models

A provider can retire a post-processing model. HyperWhisper then removes the model from the picker and maps it to a current model. Your saved modes continue to work. The app applies the map when it reads the mode, thus you do not edit the mode yourself. macOS and Windows use the same map. The replacement is always a model of the same provider, except for the local Gemma models. HyperWhisper does not move your mode to a different provider, and it does not change your API key.

​

Local LLM post-processing

After you download them, the local Gemma 4 models clean and format the transcript text offline. Your text never leaves your device. A bundled llama.cpp server runs the local LLM, and this server starts automatically when the mode needs it. Platform availability:

​

Rating scale

Every model in the library — speech-to-text and post-processing — shows a Speed bar and an Accuracy bar. Each bar has a rating from 1 to 5. The numbers come from an internal benchmark suite that runs over real recordings (results in benchmarks/results/). The Model Library sorts models by the sum of Speed + Accuracy, from high to low. Thus the most balanced models are at the top. If one rating is more important to you than the other, scroll past the top recommendations. There you can find a model that gives priority to speed or to quality.

​

Find a model in the library

The library holds every model: cloud and on-device, speech-to-text and post-processing. Use the search box and the filters at the top of Model Library to make the list shorter. Both platforms have the same set. The language filter applies only to the voice models. The post-processing models always stay in the list, because they work on text and not on audio. When you select a language, a count shows below the filters: for example 12 of 20 support Spanish. Click Show all to clear the language filter. By default the list gives the voice models first, then the post-processing models, and then the local LLM models. Each of these two first groups is in the rating order that the Rating scale section describes. To sort the whole list instead, click a column header: name, type, rating, or location. Click the same header again to change the direction.

​

How to use on-device models

​

Download and storage

Open Model Library in the app. Click Download on a model. The circular indicator shows the progress. To stop the download, click the × button. Downloaded models stay on disk until you remove them. Apple Speech is built into macOS and needs no download.

​

GPU vs CPU

Local engines use your GPU when one is available. If you have no dedicated GPU, or if the VRAM is not sufficient, the engine falls back to the CPU automatically. The model still runs, but more slowly.

​

Remove models

To make disk space available, click the trash icon next to a downloaded model in Model Library. The app removes the file immediately. You can download the model again at any time.

​

Which model to choose

​

Boost accuracy on any model

• Custom vocabulary — add product names, technical terms, and the names of colleagues. This gives the largest improvement for technical or professional use. (Support is different for each model. Apple Speech and Whisper support it locally. Most cloud providers support it, but some do not.)
• Low-noise environment — background noise makes every model less accurate. See Best Practices.
• Natural pace — speech that is too fast or too slow decreases accuracy.

​

More information