Nvidia is releasing Nemotron 3.5 Transcribe, an open speech model for transcription, translation, and speaker attribution. Although it primarily targets application makers, it also shows users what local Artificial Intelligence (AI) can now accomplish. At the same time, tools such as Anything LLM and more capable computers are enabling personal assistants whose data does not necessarily have to be processed in the cloud.
What open local AI can do
Local AI runs on your own computer or other hardware under your immediate control. Once downloaded, a model does not necessarily need an internet connection for its core tasks, and your input does not automatically have to travel to a US provider’s cloud. The t3n guide to local AI hardware therefore identifies data control, offline use, and the potential absence of a subscription as key benefits.
“Open” can mean different things in this context. An open-weight model lets you download its trained model values and run them on suitable hardware, but that is not automatically the same as fully open-source software. Nvidia describes Nemotron 3.5 Transcribe as a free open-weight model, while Granary v2 is offered as an open speech dataset with tools for data preparation and adaptation.
Local models are not limited to audio recordings. Depending on the model, they can help you draft text, create images, or work with files stored on your device. When paired with agentic software, meaning programs that carry out several steps autonomously, a model can even organize a computer’s download folder, according to t3n.
Another approach is Anything LLM as a local assistant. According to the hands-on review, the tool combines a knowledge store, agentic features, and web search. A Large Language Model (LLM), a model trained to process and generate language, can therefore provide a local chat interface for your own material; network-based functions such as web search naturally remain disabled during strictly offline use.
What Nemotron 3.5 Transcribe provides
Nemotron 3.5 Transcribe processes recorded speech in up to 40 languages. Nvidia says it can translate 31 languages to or from English, attribute speakers in 13 languages, and handle recordings with as many as eight participants. A recording can be up to one hour long, but the model does not support livestreams, according to the main report from TechCrunch.
The practical difference from a basic dictation tool is control through text instructions. From the same meeting recording, the model can produce a transcript with speaker labels, an English translation, or a summary of the decisions. For a medical recording, users can supply medication names as guidance, although Nvidia says the model should only use them when they actually occur in the audio.
Those capabilities support two concrete use cases. A team can turn an hour-long meeting into minutes separated by speaker, while a medical professional can transcribe a dictation using supplied specialist vocabulary. The model is also supposed to answer questions about a recording, although the result will still depend on the language, audio quality, accent, and subject area.
The model contains 30 billion parameters, which are values learned during training. Only about four billion are said to be active during each processing step, allowing Nvidia to position it for operation on a single Graphics Processing Unit (GPU), a chip designed for highly parallel calculations. The Decoder also describes Granary v2 as containing 844,000 hours of audio, accompanied by tools for adapting the system to specialist fields.
Nvidia reports leading results in several comparisons, including Hugging Face’s Open ASR Leaderboard. Automatic Speech Recognition (ASR) is the conversion of spoken language into written text. These provider-reported results are useful context, but they do not replace an independent test using your own recordings.
Pros and Cons of local AI
Pros:
- Greater data control – Recordings, documents, and prompts can remain on your device instead of automatically going to a cloud service.
- Offline operation – A downloaded model can perform its local core functions without an internet connection, including when connectivity is unreliable.
- No mandatory subscription – Nemotron 3.5 Transcribe is offered free of charge, and local models do not inherently require an ongoing cloud subscription.
- Adaptable output – Text instructions and supplied terminology can produce different results from the same recording, such as minutes or a translation.
Cons:
- Demanding hardware – Not every model runs well on every computer; model size, memory, and available computing power restrict your options.
- Setup rather than instant access – Local tools require more selection and configuration than a finished chat service in a browser.
- Limited functions – Nemotron only processes recordings rather than live streams, while online search cannot work entirely offline.
- Unresolved errors – Strong leaderboard positions do not rule out incorrect words, confused speakers, or inaccurate summaries.
What you can try in practice
If you are a beginner, start with one clearly defined task rather than immediately building a complete replacement for a cloud chatbot. According to t3n, a local assistant such as Anything LLM can work with an integrated knowledge store. This lets you check whether your existing hardware runs a suitable model smoothly and whether the quality is sufficient for your own texts or documents.
For speech, a recorded conversation with a small number of participants is a sensible first test. Compare the transcript with the audio and pay particular attention to names, numbers, specialist terms, and speaker labels. Human review remains necessary for confidential meetings or medical dictation even when processing takes place locally.
Advanced users can define Nemotron’s output more precisely through text instructions or adapt the model to specialist terminology and accents. The Nvidia article on Hugging Face describes Granary v2, a reusable data pipeline, and fine-tuning tools, which adjust an existing model through additional targeted training. That work is aimed more at technically experienced users and organizations than at someone who only transcribes an occasional conversation.
Local assistants can also be expanded further. Anything LLM combines a knowledge store with agentic features and web search, while other agentic programs can perform tasks on the computer. Once an assistant accesses the web or an external service, however, “local” no longer automatically means that all data remains offline.
What hardware and cost limit
The main obstacle to local AI is often not the model but the computer. Nvidia lists systems including DGX Spark, RTX Spark, Jetson AGX Thor, and Blackwell hardware for Nemotron 3.5 Transcribe. TechCrunch says the 30-billion-parameter model is intended to run on one chip, while Nvidia’s Hugging Face article says quantized versions are still coming; quantization reduces a model’s memory and computing requirements.
Those statements do not demonstrate that Nemotron will run comfortably on every ordinary laptop. The t3n hardware guide emphasizes that model and hardware suitability must be considered together and that available budget plays a major role. A smaller model that performs reasonably on your existing computer may be more useful than the most capable available model if the latter keeps the machine permanently overloaded.
Apple is now explicitly promoting local AI as a reason to buy new hardware. According to the company, the new 14-inch MacBook Pro with an M6 chip delivers up to 25 percent more performance than its predecessor for local AI tasks such as processing LLM prompts. The Swiss report on the MacBook Pro announcement gives a starting price of 1,899 Swiss francs for 16 gigabytes of memory and a shipping date of November 6.
For Switzerland, that is a specific but costly hardware option. The Nemotron sources mention support for 40 languages but do not provide a complete list, so they do not support a reliable claim about Swiss German or the quality of Switzerland’s national languages. Local processing can improve control over sensitive recordings, but it does not eliminate the need to examine the software, model license, and any enabled online features.
Local and open AI is already useful for clearly bounded tasks: transcription, work with personal documents, and simple computer actions do not necessarily have to happen in the cloud. In return, you accept greater dependence on hardware, more setup, and more responsibility for quality control. The unresolved risk is that provider benchmarks and broad language claims may look considerably stronger than a difficult real-world conversation involving several voices, accents, and specialist terms.
Sources
- Voice AI in vogue: Nvidia releases new open transcription model – TechCrunch, 2026-10-28
- Nemotron 3.5 Transcribe: Nvidias offenes Modell vereint Transkription, Übersetzung und Sprechererkennung – The Decoder, 2026-10-28
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt – Hugging Face, 2026-10-28
- Anything LLM im Praxistest: So baust du dir deinen eigenen ChatGPT-Ersatz – komplett offline – t3n, 2026-10-28
- Apple setzt mit neuem Macbook Pro auf lokale KI – Netzwoche, 2026-10-28
- Reicht schon ein Macbook mit M1? Diese Hardware brauchst du für lokale KI – t3n, 2026-10-28
Image: Jeremy Enns via Pexels