RTX Spark Laptops Start at $2,599 as Nvidia Opens Speech AI

NVIDIA said on 28 October 2026 that its RTX Spark laptops are available now, “starting at $2,599, with 128GB configurations on shelves”, according to a post on its corporate blog. The laptops come from ASUS, Dell, HP, Lenovo, Microsoft and MSI. In the same post it introduced Nemotron 3.5 Transcribe, an open-weight speech model that “turns recorded conversations into useful information”, and said its DGX Spark 64GB system “is now available to purchase from manufacturer partners”.

The speech model’s Hugging Face card says it covers 40 languages, handles recordings up to one hour, and is “ready for commercial or non-commercial use” under the OpenMDW-1.1 license. It lists 34 billion parameters in total, with about 4.3 billion active, and gives one H100 80 GB GPU as the deployment target. On the card’s own benchmark table, the model posts a mean word error rate of 4.62% on a seven-language multilingual leaderboard, against 6.49% for the next-best open-weight model Nvidia lists, Qwen3-ASR-1.7B. TechCrunch reported the release the same day, along with version 2 of Granary, Nvidia’s open speech dataset.

Short explainer · Local AI brief

A vertical news Short narrated from the published headlines and opening lines of this piece and three related FourWeekMBA pieces on on-device and open AI models.

Business Pill · Open weights

An evergreen Business Engineer explainer on open-weight models: renting a model through an API or running it yourself. It explains the general idea only and does not discuss Nvidia’s speech models.

The Rest of Nvidia’s Local-AI Launch

• RTX Spark laptops. NVIDIA’s blog says they “combine an NVIDIA Blackwell RTX GPU and an NVIDIA Grace CPU in a single superchip with up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance”. It says they are as thin as 14mm and as light as three pounds, and are available from ASUS, Dell, HP, Lenovo, Microsoft and MSI starting at $2,599.
• DGX Spark 64GB. The blog says the system is powered by the GB10 Grace Blackwell Superchip and “combines 64GB of unified memory with DGX OS and the NVIDIA AI software stack for agent development, inference and fine-tuning”. The post does not give its price.
• Nemotron 3.5 Bolt E4B 4GB. The blog says it is now available and enables “natural conversations and multistep agent tasks with as little as 4GB of memory”, powering local voice assistants, game characters and embedded interfaces. It adds that Nemotron 3.5 Bolt 74B-A9B 24GB is designed to run local agents on RTX Spark.
• Software on the laptops. The blog lists new resource management controls in the NVIDIA App to split GPU resources between foreground apps and background agents, and day-zero support for Liquid AI’s LFM3 models.

What Was Posted on X

Artificial Analysis posted that Nemotron 3.5 Transcribe ties for #1 among open-weight speech-to-text models with 2.8% WER, tied with Mistral’s Voxtral Small, using 4B active parameters out of 30B total by its count.

What Nvidia Released for Speech

• One model for several speech jobs. Nvidia’s blog says that “instead of needing multiple models for transcription, speaker labels and translation, developers can use one promptable model that handles all three, plus questions about the audio.”
• Languages. The card lists 40 languages. Its language matrix marks translation to and from English for 32 rows, including English itself, which by our count leaves 31 other languages, the figure TechCrunch reported. Nvidia’s blog says the 40 languages include seven Indic languages.
• Size. The card gives 34B total parameters and 4.3B active, in a hybrid Mamba-Transformer mixture-of-experts design, released in BF16, FP8 and NVFP4 precision. TechCrunch described the model as having 30 billion parameters; the model’s name on Hugging Face is NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.
• Recorded audio only. The card calls it an “offline speech and audio model”. TechCrunch noted that Transcribe works with recorded audio, not live streams.
• Where it runs. Nvidia’s blog lists support on RTX Spark and DGX Spark systems, Jetson AGX Thor devices and GeForce RTX 5090 GPUs, with the weights on Hugging Face from 28 October.

What Nvidia’s Benchmarks Show

All figures below come from the summary results table on Nvidia’s model card, with results captured on 12 October 2026. Lower error rates are better.

• English. A mean word error rate of 4.71% across nine public sets on the Open ASR Leaderboard, against 4.94% for Qwen3-ASR-1.7B.
• Long recordings. 10.98% on the leaderboard’s long-form sets, against 13.37% for Nvidia’s own Parakeet-TDT-0.6B-v3.
• Several languages. 4.62% on the multilingual leaderboard’s seven languages, against 6.49% for Qwen3-ASR-1.7B.
• Meetings with several speakers. A mean cpWER of 15.80 across AMI, CALLHOME, Mixer 6 and NOTSOFAR-1, against 16.83 for DiCoW v3.3 + DiariZen, which the card lists as the next-best published result.

Nvidia’s blog adds that the model “ranks No. 1 on the Hugging Face Open ASR and No. 2 among open weight models on Artificial Analysis”, with “approximately 4.3 billion active parameters, 80% fewer than the No. 1 open weight model”. By our arithmetic from the card, the gaps over the next-best open-weight models are 0.23, 2.39 and 1.87 percentage points on the English, long-form and multilingual tables.

Artificial Analysis, an independent benchmarking firm, said on X on 28 October that the model “scores 2.8% on the AA-WER Index, tied with Mistral’s Voxtral Small for #1 among open-weight models”, and that it matches Voxtral Small’s accuracy “with 6x fewer active parameters (4B vs. 24B)”. Artificial Analysis gives the model as 30B total and 4B active parameters; Nvidia’s card lists 34B total and 4.3B active, including the audio encoder.

Speed and Cost to Run

The card says that, with the optimized implementation in Nvidia’s Transcribe NIM, one RTX PRO 6000 Blackwell Server Edition GPU transcribes up to 2,238 seconds of audio per second in NVFP4 with speculative decoding, and returns the transcript of a 30-second recording in 0.13 seconds. It also says those serving optimizations “are not yet part of the open-source release”, and that the open-source vLLM integration shipped on release day is “expected to be roughly two times slower”.

Connecting the Coverage

Our Apple M6 piece reported Apple’s claim that the new iMac with M6, starting at $1,499 in the US, delivers up to 5x faster on-device AI performance than M4. Our LFM3 piece covered Liquid AI’s family of six open-weight models of about 2B parameters each, which NVIDIA now says run on RTX Spark. What’s new: NVIDIA’s blog lists Nemotron 3.5 Transcribe and Liquid AI’s LFM3 among the models available on RTX Spark, in the same post that gives the laptops’ $2,599 starting price.

Our Voice Arena piece reported that ElevenLabs models ranked first on all five new language leaderboards Voice Arena added on 23 October. Our Block piece covered Dessa, a speech model Block’s research lab says it trained only on recordings of human conversation. What’s new: Nvidia is giving away the weights of a transcription model and ranking it against open rivals on its own card.

Today’s Mistral piece covered Mistral serving DeepSeek V4.1 Flash, an open-weight model, on its API, and our EmbeddingGemma 2 piece covered Google’s technical report on a 744-million-parameter open embedding model. Our Nemotron olympiad piece reported that NVIDIA said systems built on its Nemotron 3 models reached gold-medal level at the 2026 International Olympiad in Informatics.

The Structural Read

As we read it, the card’s tables put most of the gain in long recordings and in languages other than English: the English gap is 0.23 points, against 2.39 on the long-form sets and 1.87 on the multilingual table, by our arithmetic.

Nvidia’s blog lists RTX Spark and DGX Spark systems, Jetson AGX Thor devices and GeForce RTX 5090 GPUs as supported hardware; by our reading, the release pairs free weights with a list of places to run them that is all Nvidia hardware.

The card says its fastest serving figures rely on optimizations “not yet part of the open-source release”, and that the open-source vLLM integration is expected to be roughly two times slower.

Three Implications

FOR DEVELOPERS The card lists one H100 80 GB GPU as the deployment target and releases the weights in BF16, FP8 and NVFP4 precision.

FOR TRANSCRIPTION VENDORS Nvidia’s card compares the model with open-weight rivals on error rate, not on price; the weights are free to download under the OpenMDW-1.1 license.

FOR NVIDIA Its blog lists support on RTX Spark and DGX Spark systems, Jetson AGX Thor devices and GeForce RTX 5090 GPUs.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework puts it as “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, Nvidia opening the weights of a speech model while listing its own chips as the places to run it fits that pattern. The framework was published in March 2026 and does not discuss Nvidia’s speech models.

What Is Not Established

The benchmark figures are Nvidia’s own, from its model card and blog, and compare against the models Nvidia chose to list; we did not run the model. The faster serving figures depend on an implementation Nvidia says is not yet open source. Nvidia did not publish a price, because the weights are free to download; the cost of running the model depends on the hardware used.

We could not read the Granary v2 dataset card, so this piece gives no size or language count for it. NVIDIA’s post gives a starting price for RTX Spark laptops but no price for DGX Spark 64GB or for specific laptop models. We did not contact Nvidia. Nothing here is investment advice.

The Bottom Line

NVIDIA said its RTX Spark laptops are available now starting at $2,599 and its DGX Spark 64GB system is on sale through partners, and it released Nemotron 3.5 Transcribe as an open-weight speech model for commercial and non-commercial use, covering 40 languages with about 4.3 billion of its 34 billion parameters active. On Nvidia’s own tables, the speech model’s error rates are lower than the next-best open-weight models it lists on English, long-form and multilingual transcription.