{"id":21777,"date":"2026-11-22T01:16:36","date_gmt":"2026-11-21T16:16:36","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/22\/ggml-org-hibiki-2b-gguf-2\/"},"modified":"2026-11-22T02:15:19","modified_gmt":"2026-11-21T17:15:19","slug":"ggml-org-hibiki-2b-gguf-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/22\/ggml-org-hibiki-2b-gguf-2\/","title":{"rendered":"hibiki-2b-GGUF Real-Time Audio Translation 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-2b-GGUF\">ggml-org\/hibiki-2b-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-2b-GGUF&#8221;, a GGUF format converted version of the real-time speech translation model developed by Kyutai.<\/p>\n<p>This model is an audio-to-audio and speech-to-text streaming translation model that accepts French speech as input and outputs English speech and text in real time. Unlike traditional offline translation, it does not require waiting for the end of an utterance, processing inputs continuously while generating target language speech.<\/p>\n<p>This GGUF version is provided exclusively for full-duplex operation targeting the real-time endpoint of <code>llama-server<\/code>, supporting processing that inputs and outputs audio every 80 milliseconds. Note that this model was converted using the automatic conversion tool of <code>llama-server<\/code>.<\/p>\n<h2>Specifications<\/h2>\n<p>The main specifications regarding this model and the base model &#8220;kyutai\/hibiki-2b-pytorch-bf16&#8221; are as follows:<\/p>\n<ul>\n<li>Parameter count: 2.7B (based on base model information)<\/li>\n<li>Architecture: Decoder-only hierarchical Transformer. Utilizes Moshi&#8217;s multi-stream architecture, combining and processing input and output streams<\/li>\n<li>Supported languages: French-to-English translation<\/li>\n<li>Output specifications: Generates text and audio tokens at a 12.5Hz frame rate. Generated audio bitrate is 2.2kbps<\/li>\n<li>Supported input length: Supports sequences up to 120 seconds<\/li>\n<li>Sampling method: Employs simple temperature sampling<\/li>\n<li>Audio codec: Uses the Mimi codec (12Hz, semantic and acoustic audio tokens at 1.1kbps bitrate)<\/li>\n<li>Control parameters: Supports Classifier-Free Guidance (CFG) coefficient adjustment (higher coefficients tend to increase similarity to the original speaker&#8217;s voice, while excessively high values tend to degrade translation quality)<\/li>\n<li>License conditions: CC-BY-4.0 (CC-BY in base model notation). Proper attribution is required when using<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>The evaluation of quality and functional features published in the model card of the original model &#8220;kyutai\/hibiki-2b-pytorch-bf16&#8221; is as follows. Note that the public documentation does not include specific benchmark score comparison tables.<\/p>\n<p>According to explanations by Kyutai, the developers of the base model, Hibiki is characterized by its ability to perform accurate translations in real time and in chunks while properly accumulating context, unlike traditional translation models that wait for an utterance to complete. It is also explained to have high robustness against noisy audio environments.<\/p>\n<p>In terms of processing, because it adopts simple temperature sampling, it is said to have high affinity with batch processing, unlike models requiring complex inference policies. During generation, not only audio output but also time-stamped text translations are continuously output in parallel.<\/p>\n<p>Furthermore, it is reported that the reproducibility of voice transfer in target audio generation can be adjusted by changing the Classifier-Free Guidance (CFG) coefficient. Setting a larger CFG coefficient can increase voice similarity with the input audio, but setting it too large may compromise translation accuracy.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>According to public documents for the original model &#8220;kyutai\/hibiki-2b-pytorch-bf16&#8221;, this model is designed specifically for real-time streaming speech translation (simultaneous interpretation) from French to English. Because it translates while sequentially accumulating context without waiting for the end of an utterance, it is suitable for integration into real-time conversation and interpretation applications.<\/p>\n<p>Specific application scenarios and characteristic features are as follows:<\/p>\n<ul>\n<li><strong>Real-time speech translation and text generation<\/strong>: Receives French speech input and simultaneously outputs English streaming audio and time-stamped translated text.<\/li>\n<li><strong>Robustness to noise<\/strong>: Features robustness that allows it to operate stably even against audio in noisy environments.<\/li>\n<li><strong>Voice Transfer function<\/strong>: Can generate translated audio while reflecting the voice of the input audio speaker. The strength of reproducibility can be adjusted with the CFG (Classifier-Free Guidance) coefficient.<\/li>\n<li><strong>Streaming processing of long audio<\/strong>: Supports audio sequence inputs of up to 120 seconds.<\/li>\n<li><strong>Support for batch processing<\/strong>: Because it uses simple temperature sampling rather than complex inference policies, it is also suitable for batch processing that handles numerous input sequences together.<\/li>\n<\/ul>\n<p>As background on the training data, Wikipedia, Stack Exchange, academic papers, and Common Crawl (derived from the Helium model) are used for text data, while 7 million hours of English and 450,000 hours of French pre-training data, approximately 40,000 hours of synthetic translation data, and a 900-hour fine-tuning dataset are utilized for audio data.<\/p>\n<p>As an application of components, the &#8220;Mimi&#8221; codec used within the model (12Hz, semantic and acoustic audio tokens at 1.1kpbs) can also be used standalone to develop speech language models and Text-to-Speech (TTS) systems.<\/p>\n<p>Note that the language pair currently supported by this model is French to English only. Fine-tuning is required to apply it to other language pairs. Additionally, impersonation of others and malicious use are not intended.<\/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>Q8_0<\/td>\n<td>2.2GB<\/td>\n<td>2.6GB<\/td>\n<\/tr>\n<tr>\n<td>8GB (RTX 4060 \/ 3060 Ti, etc.)<\/td>\n<td>BF16<\/td>\n<td>4.0GB<\/td>\n<td>4.9GB<\/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-2b-GGUF\/tree\/main\">ggml-org\/hibiki-2b-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-2b-BF16.gguf<\/code><\/td>\n<td>4.35GB<\/td>\n<\/tr>\n<tr>\n<td><code>hibiki-2b-Q8_0.gguf<\/code><\/td>\n<td>2.31GB<\/td>\n<\/tr>\n<tr>\n<td><code>mmproj-hibiki-2b-BF16.gguf<\/code><\/td>\n<td>1.59GB<\/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-2b-GGUF&#8221; is distributed in GGUF format at the Hugging Face <code>ggml-org\/hibiki-2b-GGUF<\/code> repository. Since a gate (<code>gated<\/code>) requiring agreement to terms of use is not set for distribution, anyone can directly access and obtain it.<\/p>\n<p>For usage, <code>llama-server<\/code> from the <code>llama.cpp<\/code> project is used. This GGUF version is designed exclusively for full-duplex operation and can only be used at the real-time endpoint of <code>llama-server<\/code> which processes audio input and output every 80 milliseconds. Please note that it is not compatible with <code>llama-tts<\/code> or <code>\/tts<\/code> endpoints.<\/p>\n<p>You can launch the model via <code>llama-server<\/code> by running the following command:<\/p>\n<pre><code class=\"language-bash\">llama-server -hf ggml-org\/hibiki-2b-GGUF -c 1500\n<\/code><\/pre>\n<p>Details on server operation methods and settings can be found in the <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/tools\/server\/README.md#full-duplex\">llama.cpp full-duplex documentation<\/a> and <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/docs\/multimodal\/tts.md#hibiki\">explanatory documentation on Hibiki<\/a>. Official code and demos for the base model are also available at <a href=\"https:\/\/github.com\/kyutai-labs\/hibiki\">Kyutai&#8217;s Hibiki repository<\/a> and the <a href=\"https:\/\/arxiv.org\/abs\/2502.03382\">paper<\/a>.<\/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\/run-hibiki-1b-gguf-real-time-audio-translation-locally\/\">hibiki-1b-GGUF Audio Conversion 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\/19\/lfm3-1-7b-gguf-release\/\">LFM3-1.7B-GGUF Text Generation Model: Our Test Answers, 4GB+ VRAM<\/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>What Q8_0 mean and where to get this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-gguf-en\/\">GGUF format guide and models<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">Quantization and model-format glossary<\/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-2b-GGUF\">https:\/\/huggingface.co\/ggml-org\/hibiki-2b-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/kyutai\/hibiki-2b-pytorch-bf16\">https:\/\/huggingface.co\/kyutai\/hibiki-2b-pytorch-bf16<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/tools\/server\/README.md#full-duplex\">https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/tools\/server\/README.md#full-duplex<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/docs\/multimodal\/tts.md#hibiki\">https:\/\/github.com\/ggml-org\/llama.cpp\/blob\/master\/docs\/multimodal\/tts.md#hibiki<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/kyutai-labs\/hibiki\">https:\/\/github.com\/kyutai-labs\/hibiki<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2502.03382\">https:\/\/arxiv.org\/abs\/2502.03382<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>ggml-org has released ggml-org\/hibiki-2b-GGUF, a GGUF version of Kyutai&#8217;s real-time French-to-English speech translation model for llama-server.<\/p>\n","protected":false},"author":1,"featured_media":21776,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[161,1970,163,5533,5535,1073,1547,5049,5537],"class_list":["post-21777","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-ggml-en","tag-ggml-org-en","tag-gguf-en","tag-hibiki-2b-en","tag-hibiki-pytorch-en","tag-llama-cpp-en","tag-verified","tag--en"],"lang":"en","translations":{"en":21777,"ja":21775},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21777","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=21777"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21777\/revisions"}],"predecessor-version":[{"id":22019,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21777\/revisions\/22019"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/21776"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=21777"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=21777"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=21777"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}