{"id":16173,"date":"2026-11-17T11:04:35","date_gmt":"2026-11-17T02:04:35","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=16173"},"modified":"2026-11-17T11:04:36","modified_gmt":"2026-11-17T02:04:36","slug":"haru-4b-speech-model","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/haru-4b-speech-model\/","title":{"rendered":"Haru-4B Speech-to-Speech Model: 12GB+ VRAM, File List"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/higurashi-ai\/Haru-4B\">higurashi-ai\/Haru-4B<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-12<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p><code>higurashi-ai\/Haru-4B<\/code>, released by higurashi-ai, is a speech-to-speech conversational model that takes user audio input and outputs both the model&#8217;s own voice and text in real time. According to the model card, it is a foundation model for low-latency, multilingual full-duplex continuous voice conversations, allowing users to listen and speak simultaneously like on a telephone. It is a pretrained base model released for research purposes.<\/p>\n<h2>Specifications<\/h2>\n<ul>\n<li>Architecture: The language model uses a Qwen3.5-4B text backbone (32 layers, hidden size 2560).<\/li>\n<li>Frame Rate: 7.5 Hz. It receives one frame of user audio every 133 ms and generates one frame of model audio and text. Because user audio, model audio, and model text are added to a single position in the language model, listening and speaking processes do not wait for each other.<\/li>\n<li>Audio Representation: Uses 64-dimensional continuous latent representations per frame, with no discrete speech tokens or separate audio transformers.<\/li>\n<li>Audio Head: Uses a 4-layer MLP with MeanFlow sampling, generating in 1 step. The guidance scale is 1 to 2 (default 1.5).<\/li>\n<li>Audio Codec: VibeVoice acoustic tokenizer (causal, 24 kHz, 3200 samples per latent). Both input and output handle 3200 samples at 24 kHz.<\/li>\n<li>Supported Languages: 9 languages, including English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean.<\/li>\n<li>Character and Voice Control: Characters are specified via a text character card, and voices are specified using a short reference audio clip.<\/li>\n<li>Reference Audio Conditions: Voice clips are not bundled and must be provided by the user. They should be clean single-speaker audio lasting from 3 seconds to 3 minutes. Users are requested to use only voices they have the right to use.<\/li>\n<li>License Conditions: The documentation contains no statements regarding commercial use permissions or restrictions. However, the model card describes the release as being for research and draws attention to the fact that it can mimic reference voices.<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>There are no benchmarks or comparison tables with other models in the model card. Regarding quality, the publishers state as limitations that while it enables natural conversation, it is not reliable as a factual information source and does not follow instructions like an assistant.<\/p>\n<p>For low latency, the use of wired headphones and a wired microphone is recommended. Bluetooth devices are reported to add 100 ms or more of latency. A technical blog has been published, but its contents have not been verified here as it was not included in the source materials.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>According to the model card, the primary use case is full-duplex continuous voice conversation where users can &#8220;listen and speak simultaneously&#8221; like on a telephone. Because it returns responses while receiving user audio frame by frame, unlike turn-based voice assistants, it is suited for research and experimentation in real-time banter. The tags include <code>speech-to-speech<\/code> and <code>audio-to-audio<\/code>.<\/p>\n<p>Specific use cases include:<\/p>\n<ul>\n<li>Building conversation demos with characters set via character cards and voices specified by reference audio<\/li>\n<li>Conversation experiments in 9 languages: English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean<\/li>\n<li>A research baseline for low-latency voice conversations<\/li>\n<\/ul>\n<p>There are also points to keep in mind. The publishers state that this is a pretrained base model, is not reliable as a factual source, and does not follow instructions like an assistant. It is not intended as a commercial question-and-answering bot. Additionally, because it can mimic reference voices, users are asked to use only voices they have the right to use.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 5.3B parameters<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>12GB (RTX 4070 \/ 3060 12GB, etc.)<\/td>\n<td>BF16<\/td>\n<td>9.8GB<\/td>\n<td>11.8GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><strong>Inference engine support<\/strong> (architecture name matched against each project&#8217;s own model registry in its source code, checked 2026-11-17): llama.cpp: not registered, vLLM: not registered, MLX (mlx-lm): not registered. &#8220;Not registered&#8221; means the name is absent from that registry today, not that the model cannot run.<\/p>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:hardware --><\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/higurashi-ai\/Haru-4B\/tree\/main\">higurashi-ai\/Haru-4B<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>Size<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>model.safetensors<\/code><\/td>\n<td>10.54GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>The distribution format is safetensors, and the repository is <code>higurashi-ai\/Haru-4B<\/code>. No gating (agreement to terms of service) is set, and it can be obtained without consent. The included files are as follows:<\/p>\n<ul>\n<li><code>model.safetensors<\/code>: All weights (bfloat16)<\/li>\n<li><code>config.json<\/code>: Model configuration<\/li>\n<li><code>tokenizer.json<\/code>: Text tokenizer<\/li>\n<\/ul>\n<p>Official code is available on GitHub at <a href=\"https:\/\/github.com\/higurashi-ai\/Haru\">higurashi-ai\/Haru<\/a>. It is described as including a PyTorch implementation, a web demo, an MLX implementation with a native macOS app, and a plugin for serving with vLLM-Omni. The steps to run it on an NVIDIA GPU are as follows:<\/p>\n<pre><code class=\"language-bash\">git clone https:\/\/github.com\/higurashi-ai\/Haru &amp;&amp; cd Haru\npip install -e.\/haru\npython -m haru.server --hf-repo higurashi-ai\/Haru-4B\n<\/code><\/pre>\n<p>After launching, open the web demo, select a character, provide a voice, and start speaking. The model card suggests budgeting about 12 GB of memory on an NVIDIA GPU.<\/p>\n<p>When using from Python, load the model with <code>HaruModel.from_pretrained(\"higurashi-ai\/Haru-4B\")<\/code>. Prepare reference audio with <code>prepare_reference(load_audio(\"voice.wav\"))<\/code>, and pass the character card and reference audio to <code>DuplexSession<\/code>. Then obtain the first 133 ms using <code>session.start()<\/code> and repeat <code>session.step(user_audio)<\/code>. Inputs and outputs are 3200 samples at 24 kHz. The system prompt is a fixed-layout character card, and instructions on how to create one are noted in the repository&#8217;s README.<\/p>\n<p>For low-latency usage, wired headphones and a wired microphone are recommended.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/24\/tev1-4b-experimental-released\/\">Tev1-4B-experimental Text Generation Model: 4GB+ VRAM, GGUF Builds<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Find models by VRAM<\/strong> (This model runs from the 12GB tier) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-12gb-en\/\">Other models that run on a 12GB GPU<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-audio\">Other audio models<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/higurashi-ai\/Haru-4B\">higurashi-ai\/Haru-4B(Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/higurashi-ai\/Haru\">higurashi-ai\/Haru(GitHub)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Explore Haru-4B by higurashi-ai, a low-latency, multilingual full-duplex speech-to-speech conversational base model for research.<\/p>\n","protected":false},"author":1,"featured_media":16172,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[5039,5041,5043,5045,996,209,5047,1547,5049],"class_list":["post-16173","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-audio-to-audio-en","tag-haru-en","tag-higurashi-ai-en","tag-higurashi-ai-haru-4b-en","tag-qwen3-5-en","tag-safetensors-en","tag-speech-to-speech-en","tag-verified","tag--en"],"lang":"en","translations":{"en":16173,"ja":16171},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16173","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=16173"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16173\/revisions"}],"predecessor-version":[{"id":17192,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16173\/revisions\/17192"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/16172"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=16173"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=16173"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=16173"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}