March 2025 Summaries

Modal announced price reductions of up to 65% for GPU, CPU, and memory resources, alongside Team plan support for scaling to 50 concurrent GPUs. The Modal SDK is now available through conda-forge beginning with version 0.73.46, responding to demand from scientific computing users. As the company prepares for version 1.0 of its client, it has released a migration guide to help users address deprecations and SDK changes, including the transition of experimental buffer containers into the standard `buffer_containers` decorator option. Other updates include an LTX-Video image-animation example, a blog post on measuring and improving GPU utilization, and an open-source GPU glossary accepting community contributions. Modal also introduced a new logo and began a San Francisco billboard campaign with prizes for people who share photos of the ads.

The text-to-speech (TTS) landscape is rapidly changing, with new state-of-the-art models launching every month. Developers and businesses are seeking powerful, flexible, and cost-effective TTS options, and several open-source libraries have emerged to address this need. Spark-TTS is a 500 million parameter model that supports zero-shot voice cloning, bi-lingual speech synthesis, and adjustable voice attributes. Kokoro is a super-small TTS model with 82M parameters, offering fast deployment and high-quality audio at a lower cost. Fish Speech v1.5 features low CER/WER, fast latency, and support for multiple languages, but its license restricts commercial use. xtts-v2 supports 13 languages and expressive speech synthesis, while StyleTTS produces exceptionally natural-sounding English speech with a permissive license. OpenVoice v2 offers instant voice cloning capabilities, but with limited language support compared to MeloTTS. VITS is a lightweight model suitable for on-device use cases like article reading or language practice. These open-source TTS libraries offer alternatives to commercial solutions and can be used in conjunction with Modal for real-time applications.

The landscape of open-source large language models (LLMs) has become increasingly crowded, making it challenging for developers to navigate the options and find the best model for their specific use case. To help with this, we'll explore some of the top open-source LLMs, evaluating them based on factors such as performance, size, ease of use, and suitability for various tasks like writing, coding, and fine-tuning. The models have been compared in terms of quality, speed, cost, and other metrics to provide a comprehensive overview of their capabilities. We also discuss the differences between instruct and non-instruct versions, 7B and 70B parameter counts, and how to serve these models blazingly fast using platforms like Modal.