TL;DR
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Cactus Compute announced Whistle, a 16.9 MB speech-recognition model designed to run on-device using the company’s C++ CPU engine. The company reports support for seven languages and publishes benchmark comparisons, but those results depend on the selected datasets, runtimes and test setup.
Cactus Compute announced Whistle, a 16.9 MB speech-recognition model that the company says can transcribe audio locally on a CPU in seven languages. The release is aimed at devices including phones, wearables, robots, smart-home products and cars, where keeping speech processing on-device may reduce reliance on an internet connection or remote service.
According to the company’s October 2 announcement, Whistle processes 16 kHz mono audio clips up to 30 seconds in English, German, French, Spanish, Italian, Dutch and Polish. It detects the language unless the user specifies one. Cactus Compute says the model runs without dependencies in its C++ engine, the same engine used for its Needle model, and that audio in the browser demonstration stays on the device after the model is downloaded.
The release describes three functions: transcription, word-level timestamps and speech embeddings. The timestamps include a start and end time and a probability for each word. The embedding mode returns encoder output at one row per 80 milliseconds without generating a transcript. A silence check can return an empty transcript without running the decoder’s beam search, the company says.
Cactus Compute reports that Whistle’s model file is 16.9 MB, compared with 145.3 MB for Whisper base and 41.9 MB for Moonshine tiny v2 in its comparison. In a test using 10 seconds of audio on an Apple M4 Pro CPU, it reports 11.1 milliseconds to the first token and 1,319 tokens per second during decoding. The company says it used each model’s official runtime and default settings, with Whistle configured for five beams. These are company-reported measurements, not an independently verified assessment.
Smaller Models for Local Speech
A compact model that can run without a server could make speech input more practical on devices with limited storage, intermittent connectivity or restrictions on sending audio to cloud services. The 16.9 MB file size is central to Cactus Compute’s pitch for use in mobile and embedded products. Local processing can also keep audio on a device, though actual privacy protections depend on how an application handles recordings and other data.
The release also targets developers already using Needle: Whistle is designed to load into the same CPU engine and container. Cactus Compute says audio can be transcribed and then passed to tools using one binary. That could simplify some deployments, but the announcement does not establish how much memory, battery power or integration work a real product would require. Model size and benchmark speed alone do not show how well a system performs under every device or use case.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
How Whistle Fits Cactus Compute’s Stack
Cactus Compute describes Whistle as sharing components with Needle, its existing model, rather than as a standalone architecture unrelated to that product. Its speech pipeline converts audio into log-mel features, processes those through an encoder, and uses a decoder to generate text. The company says the encoder runs all eight blocks at every selected decoder depth; developers can choose decoder depth when loading the model.
The published comparison reports mixed accuracy results rather than a win on every dataset. Cactus Compute says Whistle scores better than Whisper base on LibriSpeech test-clean and test-other, SPGISpeech, Earnings-22 and the FLEURS average, while Whisper base scores better on TED-LIUM, AMI and the MLS average. Moonshine tiny v2 is English-only, and the company notes that missing benchmark bars reflect datasets for which model authors did not publish results. It also warns that Whisper’s AMI result uses AMI-IHM, a different subset from the AMI figures reported for the other models.
““One 16.9 MB file, runs on the CPU with no dependencies.””
— Cactus Compute, in its October 2, 2026 release
Independent Tests and Device Demands
The release does not provide independent verification of the accuracy, speed or file-size comparisons. Benchmark results can vary with hardware, runtime configuration, audio duration and dataset selection. The company identifies the test hardware and some settings, but the supplied material does not establish performance across a broad range of phones, wearables or other target devices.
It is also unclear from the announcement how Whistle performs in noisy environments, with accents or vocabulary outside the reported tests, or in longer recordings beyond the stated 30-second limit per pass. Cactus Compute presents the browser demo as on-device, but does not provide a broader audit of data handling by products built with the model. The release does not specify licensing terms or provide evidence of independent evaluation.
Availability and Further Evaluation
Cactus Compute’s release includes a browser demonstration that downloads the 16.9 MB model on first use and accepts clips up to 30 seconds. Developers and prospective users can compare the model’s behavior with the company’s published benchmarks, while treating those measurements as vendor-reported results.
The next useful evidence would include independent benchmark runs, clear licensing and deployment documentation, and tests on the devices Whistle is intended to serve. Cactus Compute has not stated in the supplied announcement when further evaluations or additional language support might be available.
Key Questions
What is Whistle?
Whistle is a speech-recognition model released by Cactus Compute. The company says it transcribes audio locally using its C++ CPU engine.
Which languages does Whistle support?
Cactus Compute lists English, German, French, Spanish, Italian, Dutch and Polish. The model detects the language unless a user specifies it.
Can Whistle process audio without sending it to a server?
The company says Whistle runs on-device and that audio in its browser demonstration does not leave the device. That claim describes the company’s setup; data handling in other applications depends on their implementation.
How does Whistle compare with Whisper base?
Cactus Compute reports a smaller file size and faster first-token and decoding measurements for Whistle in its stated test. Its reported accuracy comparisons are mixed: Whistle leads on some datasets, while Whisper base leads on others. The results have not been independently verified in the supplied material.
What audio can Whistle transcribe in one pass?
The release specifies 16 kHz mono audio up to 30 seconds per pass. Cactus Compute has not described longer-clip performance in the supplied announcement.
Source: hn
![The GNU Emacs Architecture: Unlocking The Core [Pdf]](https://cables-and-networks.com/wp-content/uploads/2026/07/the-gnu-emacs-architecture-unlocking-the-core-pdf-featured-560x300.jpg)
