Want to run smooth English text-to-speech directly on your laptop or in your browser? The newly open-sourced phonon-31m-en has only 31M parameters, and after quantization, it takes up just 64.5 MB. Its official blind test results even beat the well-known Kokoro model. In this issue, we will also share a 28B reasoning model that can fit onto a single GPU, as well as a powerhouse video generation model dedicated to ComfyUI with over 780,000 downloads. We will show you how to deploy them locally in just a moment.

---

Featured This Week

1. gradium/phonon-31m-en: Ultra-Lightweight 24 kHz Local Text-to-Speech Model

• What is it: This is an ultra-lightweight text-to-speech (TTS) model licensed under Apache 2.0. With only 31M parameters, it is designed specifically for highly efficient local voice generation.
• Use Cases:
  • Independent and private local smart home voice assistants: Instantly convert system status or smart home notifications into smooth 24 kHz English speech without an internet connection, completely eliminating concerns about private conversations being uploaded to the cloud.
  • Offline games and interactive application NPCs: Leverage its streaming capabilities to play audio in real-time as it generates, completely eliminating the network latency of cloud APIs to make in-game NPC dialogue responses more instant and natural.
  • Local accessibility voice prompts: Integrate it into Home Assistant or use it as an OpenAI API-compatible local voice server to provide ultra-fast, low-latency screen reading and accessibility voice prompts for visually impaired users.
• Can you run it: The official model card indicates that the model is lightweight enough to run directly in a browser tab, on a mobile phone, or on a laptop CPU. The 8-bit quantized weight file (model.q8.gguf) requires only 64.5 MB. Available formats include GGUF and Safetensors.
• Community Traction: It has reached 112 trending score on Hugging Face, currently accumulating 39.5k downloads and 115 likes.
• How it compares: According to the blind pairwise comparison (Elo) and hard set LLM Word Error Rate (WER) benchmarks provided in the official model card, phonon-31m-en outperforms similar well-known open-source models like Kokoro (Elo 2043, WER 3.5%) and Magpie (Elo 1939, WER 28%) with an Elo rating of 2098 and a WER of 1.5%.
• How to get it: The official documentation states that it can be loaded via Python's ptts library. First run pip install ptts, then use tts = ptts.TTS("hf://gradium/phonon-31m-en", lang="en", quant="q8") for initialization and voice generation.
• Link: Hugging Face Page

---

2. KRAFTON/Raon-1-mini-preview: MoE Inference Model Supporting Ultra-Long Context and Thinking Mode

• What is it: This is a 28.8B Mixture-of-Experts (MoE) architecture model released by KRAFTON under the Apache 2.0 license. Although the total parameter count is 28.8B, it only requires 4.3B active parameters per token and features extreme KV Cache optimization.
• Use Cases:
  • Local Agents and automated development: Developers can use Agent frameworks like Hermes Agent, Pi, or OpenCode locally, calling Raon-1 mini via an OpenAI-compatible API to perform complex tool calling and automated code generation.
  • Single-GPU local complex reasoning: Perform complex local mathematical and scientific reasoning on consumer-grade graphics cards (such as a single RTX 5090) or Apple Silicon chips, utilizing its built-in "thinking mode" and long-context support of up to 262,144 tokens.
  • Ultra-fast inference output: Pair it with the officially released DFlash2 lightweight draft model for speculative decoding, achieving over 2x faster inference output on local GPUs.
• Can you run it: The official README indicates that in its MoE architecture, only 7 MLA layers retain the KV Cache (requiring only about 8 KiB per token in BF16 precision). The official release provides MXFP8 and NVFP4 quantized versions, where the NVFP4 version allows inference on a single NVIDIA RTX 5090 (32 GB VRAM). GGUF (Q4_K_M) and MLX versions are also available, offering native support for llama.cpp and MLX.
• Community Traction: It has reached 185 trending score on Hugging Face, currently with 1.4k downloads and 200 likes.
• How it compares: According to tests in the official README, in comparisons with other MoE baseline models (including Qwen3.6-35B-A3B and Gemma 4 26B-A4B), Raon-1 mini preview achieved the best results in knowledge and reasoning benchmarks such as MMLU Pro (76.5), GPQA Diamond (84.3), and IFBench (84.2).
• How to get it: The official documentation states that the model repository is currently gated. You need to request automatic access on the Hugging Face page first, and then run it using llama.cpp: ./build/bin/llama-cli -hf KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M --ctx-size 262144 --reasoning on.
• Link: Hugging Face Page

---

3. WarmBloodAban/Minimax-h3_Singularity: Ultra-Smooth Fine-Tuned Video Generation Model for ComfyUI

• What is it: This is a video generation model dedicated to ComfyUI, fine-tuned and merged based on the MiniMax-H3 base model. Licensed under Apache 2.0, it is designed specifically for engineers who love self-hosting and local creation.
• Use Cases:
  • High-action combat video generation: Creators can use Image-to-Video (I2V) workflows in local ComfyUI to transform static High Dynamic Range (HDR) images into smooth, motion-blur-free, high-speed martial arts or sword combat videos.
  • Distant face restoration in medium-to-long shots: Video post-production editors can leverage its "Distant Face Restoration" feature when generating medium-to-long shots, preventing character faces in the generated videos from distorting or collapsing structurally.
  • Dynamic assets for game art: Game artists can take advantage of its optimization for visual effects (VFX) to quickly generate dynamic assets locally, such as magic auras, particle effects, or spell-casting animations.
• Can you run it: The official release includes a V2 INT8 quantized version. To achieve high-speed generation, the author recommends pairing it with the minimax_h3_ref2v_turbo_4step_v0.1 acceleration LoRA, enabling fast inference in just 4 steps.
• Community Traction: It has reached 127 trending score on Hugging Face, accumulating a massive 783.2k downloads and 916 likes, making it highly popular among community creators.
• How it compares: According to the author's model card, this model is based on the MiniMax-H3 base model, merging multiple key fine-tuning checkpoints such as ref, fl, and b25-49, and undergoing 3 full days of precise pruning and weight optimization. Compared to the original MiniMax-H3, it significantly reduces overly oily skin and unnatural gloss, and fixes the issue in the V1 version where Audio Reference Conditioning failed to work.
• How to get it: The official documentation states that after downloading the weights, you can place them directly into the corresponding model folder in ComfyUI. It supports various workflows including Text-to-Video, Image-to-Video, Reference-to-Video, and Video-to-Video.
• Link: Hugging Face Page

---

Conclusion

Looking at the models introduced today, the trend of local AI is clearly moving toward extreme lightweighting and specialization. Whether it is a text-to-speech model that runs on just 64MB or a highly optimized 28B reasoning model that fits on a single GPU, local deployment is becoming much more accessible. If you have some spare hardware lying around, why not start playing with phonon-31m-en, which takes up the least space? See you in the next episode.