{"id":21976,"date":"2026-11-22T02:10:27","date_gmt":"2026-11-21T17:10:27","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/22\/run-hibiki-1b-gguf-real-time-audio-translation-locally\/"},"modified":"2026-11-22T02:10:27","modified_gmt":"2026-11-21T17:10:27","slug":"run-hibiki-1b-gguf-real-time-audio-translation-locally","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/22\/run-hibiki-1b-gguf-real-time-audio-translation-locally\/","title":{"rendered":"hibiki-1b-GGUF Audio Conversion Model: 4GB+ 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\/ggml-org\/hibiki-1b-GGUF\">ggml-org\/hibiki-1b-GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-22<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>cc-by-4.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-gguf-en\/\">GGUF<\/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>ggml-org has released &#8220;ggml-org\/hibiki-1b-GGUF&#8221;, which is a GGUF quantized version of the real-time speech translation model &#8220;kyutai\/hibiki-1b-pytorch-bf16&#8221;.<\/p>\n<p>This model is an audio-to-audio streaming model that takes French speech input and outputs English speech and text translation in real time. Unlike traditional translation approaches, it does not need to wait for the entire utterance to finish. Instead, it sequentially accumulates the necessary context from the incoming audio stream and instantly generates target-language speech and text in parallel.<\/p>\n<p>&#8220;ggml-org\/hibiki-1b-GGUF&#8221; is a repository converted using ggml-org&#8217;s automatic conversion tools based on the open-weight model &#8220;kyutai\/hibiki-1b-pytorch-bf16&#8221; developed by Kyutai. By utilizing the real-time endpoint (full-duplex communication) in <code>llama.cpp<\/code>&#8216;s <code>llama-server<\/code>, you can run high-speed and memory-efficient streaming speech translation locally.<\/p>\n<h2>Specifications<\/h2>\n<p>According to the published materials, the specifications for the base model &#8220;kyutai\/hibiki-1b-pytorch-bf16&#8221; and this GGUF converted version are as follows:<\/p>\n<ul>\n<li>Parameters: 1.7B (referred to as &#8220;Hibiki-M&#8221; in the research paper, a 1.7B parameter configuration)<\/li>\n<li>Architecture: Hierarchical Transformer. A decoder-only model adopting Moshi&#8217;s multi-stream structure<\/li>\n<li>Input\/Output Specifications: Input is French audio, output is English audio and time-stamped English text translation<\/li>\n<li>Token Generation Speed: 12.5Hz (continuously outputs text and audio tokens at a constant rate)<\/li>\n<li>Audio Generation Bitrate: 1.1kbps<\/li>\n<li>Supported Languages: French-to-English translation<\/li>\n<li>Audio Codec: Mimi codec (a neural audio codec integrating semantic and acoustic information at a 12Hz frame rate and 1.1kbps bitrate)<\/li>\n<li>Sampling and Control: Supports simple temperature sampling. The similarity of voice transfer can be adjusted by tuning the Classifier-Free Guidance (CFG) coefficient<\/li>\n<li>Operating Format: Full-duplex. Continuously processes audio input and output every 80ms using the <code>llama-server<\/code> real-time endpoint<\/li>\n<li>License Terms: CC-BY-4.0 (commercial and secondary use is permitted provided appropriate credit is given to the original authors)<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>According to the model card written by Kyutai, the developer of the base model &#8220;kyutai\/hibiki-1b-pytorch-bf16&#8221;, this model features exceptional real-time processing capabilities and flexible audio generation control in streaming simultaneous interpretation.<\/p>\n<p>While specific benchmark score tables are not included in the model card, the following points regarding the model&#8217;s quality and operational characteristics have been reported:<\/p>\n<p>In simultaneous interpretation performance, it autonomously determines and accumulates the minimum context required for translation without waiting for the entire utterance to complete, smoothly generating English speech in chunks. Simultaneously with this audio generation, time-stamped text translation results are also output in parallel.<\/p>\n<p>Additionally, it features a &#8220;voice transfer&#8221; function that reflects the input speaker&#8217;s voice quality in the generated audio, and its reproducibility can be controlled via the Classifier-Free Guidance (CFG) coefficient. There is an advantage that setting the CFG coefficient higher improves similarity to the input speaker&#8217;s voice quality. On the other hand, it is explained that there is a trade-off relationship where setting the coefficient excessively high degrades translation accuracy and quality.<\/p>\n<p>It also has robustness against input audio noise, and since it is trained on long sequences of up to 120 seconds, stable streaming translation is possible even during long conversations. Furthermore, because it utilizes temperature sampling which does not require complex inference policies, it has the advantage of being easily adaptable to batch processing of multiple sequences.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Here are the use cases and expected application scenarios where &#8220;ggml-org\/hibiki-1b-GGUF&#8221; and its base model &#8220;kyutai\/hibiki-1b-pytorch-bf16&#8221; demonstrate their true value.<\/p>\n<h3>Real-Time Simultaneous Interpretation and Streaming Speech Translation<\/h3>\n<p>The biggest feature of this model is its ability to process French audio input in real time and simultaneously translate it into English speech and text with minimal latency. Unlike conventional approaches that require waiting for the speaker to finish speaking the entire sentence, it sequentially accumulates context from the input audio stream and immediately outputs target-language speech continuously. Therefore, it is ideal for applications requiring low-latency processing, such as real-time dialogue systems, online meeting subtitles and voice interpretation, and live stream simultaneous interpretation.<\/p>\n<h3>Speech Conversion Reflecting Speaker Voice Quality (Voice Transfer)<\/h3>\n<p>A &#8220;voice transfer&#8221; function is integrated into the translated English speech to reflect the voice quality of the input French speaker. By adjusting the Classifier-Free Guidance (CFG) coefficient, you can control how closely it approximates the original speaker&#8217;s voice quality. This is suitable for content production and voice dialogue systems where you want to convert to another language while retaining the original speaker&#8217;s identity and atmosphere.<\/p>\n<h3>Handling Noisy Environments and Long-Length Audio<\/h3>\n<p>This model has been trained using audio sequences of up to 120 seconds and diverse audio datasets, providing high robustness against real-world input audio containing noise. It operates stably not only in short exchanges but also in voice translation during long conversations and environments with everyday background noise.<\/p>\n<h3>Parallel Interpretation via Batch Processing<\/h3>\n<p>Because it operates with simple temperature sampling without requiring complex algorithms or policies during inference, it is also suitable for batch processing multiple audio input streams simultaneously. It allows for easy integration into infrastructure configurations that run multiple interpretation sessions in parallel on the server side.<\/p>\n<h3>Application to Codecs and Other Language Models<\/h3>\n<p>The &#8220;Mimi&#8221; codec integrated into this model is a high-performance neural audio codec that integrates acoustic and semantic information at a 12Hz frame rate and a low bitrate of 1.1kbps. You can extract this component on its own to develop custom speech-language models or text-to-speech (TTS) systems. Moreover, it is suitable as a base model for fine-tuning into simultaneous interpretation models for other language pairs by leveraging Hibiki&#8217;s own architecture.<\/p>\n<p>Please note that use cases intended for impersonation or malicious spoofing of others are explicitly prohibited as Out-of-Scope in the model card.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong><\/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>4GB (laptop iGPU \/ phone class)<\/td>\n<td>BF16<\/td>\n<td>1.9GB<\/td>\n<td>2.3GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\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 GGUF.<\/p>\n<p><strong>License \u2014 <code>cc-by-4.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, but <strong>attribution is mandatory<\/strong> \u2014 omitting credit is a violation.<\/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\/ggml-org\/hibiki-1b-GGUF\/tree\/main\">ggml-org\/hibiki-1b-GGUF<\/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>hibiki-1b-BF16.gguf<\/code><\/td>\n<td>2.04GB<\/td>\n<\/tr>\n<tr>\n<td><code>hibiki-1b-Q8_0.gguf<\/code><\/td>\n<td>1.08GB<\/td>\n<\/tr>\n<tr>\n<td><code>mmproj-hibiki-1b-BF16.gguf<\/code><\/td>\n<td>1.76GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>&#8220;ggml-org\/hibiki-1b-GGUF&#8221; can be downloaded directly from the Hugging Face repository. No prior agreement to terms of service or application (gated access) is required to obtain it.<\/p>\n<p>This model is provided in GGUF format and is launched using <code>llama-server<\/code> included in <code>llama.cpp<\/code>.<\/p>\n<h3>Run Command<\/h3>\n<p>If you have your <code>llama.cpp<\/code> environment set up, running the following command will download the model and start the server:<\/p>\n<pre><code class=\"language-bash\">llama-server -hf ggml-org\/hibiki-1b-GGUF -c 1500\n<\/code><\/pre>\n<h3>Notes on Execution<\/h3>\n<p>This model is exclusively a full-duplex communication model. Therefore, it does not work through standard text-to-speech endpoints such as <code>llama-tts<\/code> or <code>\/tts<\/code>. Be sure to run it using the real-time endpoint of <code>llama-server<\/code> (full-duplex communication function that inputs and outputs audio every 80ms).<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/22\/ggml-org-hibiki-2b-gguf-2\/\">hibiki-2b-GGUF Real-Time Audio Translation Model: 4GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/22\/ggml-org-releases-voxcpm2-gguf\/\">VoxCPM2-GGUF 30-Language Speech Synthesis Model: 4GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/18\/kokoro-82m-gguf-released\/\">Kokoro-82M-GGUF Lightweight Text-to-Speech Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/higgs-audio-v3-stt-gguf\/\">higgs-audio-v3-stt-GGUF Speech-to-Text Model: 4GB+ VRAM, File List<\/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 4GB tier) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-8gb-en\/\">Other models that run on a 8GB GPU<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-gguf-en\/\">GGUF 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\/ggml-org\/hibiki-1b-GGUF\">ggml-org\/hibiki-1b-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/kyutai\/hibiki-1b-pytorch-bf16\">kyutai\/hibiki-1b-pytorch-bf16<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/kyutai-labs\/hibiki\">kyutai-labs\/hibiki (GitHub)<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2502.03382\">Hibiki Paper (arXiv:2502.03382)<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/tools\/server\/README.md#full-duplex\">llama.cpp Full-Duplex Documentation<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/docs\/multimodal\/tts.md#hibiki\">llama.cpp Hibiki TTS Documentation<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Learn about ggml-org\/hibiki-1b-GGUF, a GGUF quantized real-time audio-to-audio translation model, and how to run it locally using llama.cpp.<\/p>\n","protected":false},"author":1,"featured_media":21975,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[161,5549,163,5535,5551,1073,1547,5049,5553],"class_list":["post-21976","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-ggml-en","tag-ggml-org-hibiki-1b-gguf-en","tag-gguf-en","tag-hibiki-pytorch-en","tag-kyutai-hibiki-1b-pytorch-bf16-en","tag-llama-cpp-en","tag-verified","tag--en"],"lang":"en","translations":{"en":21976,"ja":21974},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21976","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=21976"}],"version-history":[{"count":0,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21976\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/21975"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=21976"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=21976"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=21976"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}