{"id":20062,"date":"2026-11-19T06:11:46","date_gmt":"2026-11-18T21:11:46","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/19\/rumble-2-flash-gguf-released\/"},"modified":"2026-11-19T07:21:30","modified_gmt":"2026-11-18T22:21:30","slug":"rumble-2-flash-gguf-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/19\/rumble-2-flash-gguf-released\/","title":{"rendered":"Rumble-2-Flash-GGUF 45-Language ASR Model: 4GB+ VRAM, File List"},"content":{"rendered":"<p><em>Sample outputs are available on the <a href=\"https:\/\/huggingface.co\/AtomicChat\/Rumble-2-Flash-GGUF\">model card<\/a>.<\/em><\/p>\n<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\/AtomicChat\/Rumble-2-Flash-GGUF\">AtomicChat\/Rumble-2-Flash-GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-19<\/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-gguf-en\/\">GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Unverified (not confirmed by a primary source)<\/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>It is reported that the AtomicChat team has released GGUF format files for the speech recognition model &#8220;Rumble-2-Flash&#8221;. This model is designed to take audio data as input and transcribe its contents into text (speech recognition). Note that this information has not been officially confirmed at this time.<\/p>\n<p>This model was created by the team developing &#8220;Atomic Chat&#8221;, an app for running open-weight models locally and offline. The base model is <code>Reson8\/Rumble-2-Flash<\/code>, and it has been quantized for use with <code>llama.cpp<\/code>. The model consists of an 18-layer audio encoder and a 14-layer decoder.<\/p>\n<h2>Specifications<\/h2>\n<p>The specifications listed in the documentation are as follows:<\/p>\n<ul>\n<li>Architecture: 18-layer audio encoder and 14-layer decoder<\/li>\n<li>Base model: <code>Qwen3-ASR-0.6B<\/code><\/li>\n<li>Number of parameters: 0.5B<\/li>\n<li>Supported languages: 45 languages<\/li>\n<li>License: Apache 2.0<\/li>\n<li>Distribution format: GGUF<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>The model card shows performance comparison data across different quantization levels. The evaluation uses datasets such as LibriSpeech, Multilingual LibriSpeech (MLS), and FLEURS. Performance metrics used are WER (Word Error Rate, lower is better) and CER (Character Error Rate, lower is better) for transcriptions.<\/p>\n<p>The following table shows a comparison of performance across each quantized file. Note that these figures are results measured using fixes applied to <code>llama.cpp<\/code>.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>GB on disk<\/th>\n<th>bpw<\/th>\n<th>LS clean<\/th>\n<th>LS other<\/th>\n<th>MLS DE<\/th>\n<th>MLS ES<\/th>\n<th>MLS FR<\/th>\n<th>FLEURS ZH<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>Rumble-2-Flash-BF16.gguf<\/code><\/td>\n<td>0.65<\/td>\n<td>16.0<\/td>\n<td>2.09<\/td>\n<td>4.70<\/td>\n<td>6.85<\/td>\n<td>4.57<\/td>\n<td>6.57<\/td>\n<td>10.38<\/td>\n<\/tr>\n<tr>\n<td><code>Rumble-2-Flash-Q8_0.gguf<\/code><\/td>\n<td>0.35<\/td>\n<td>8.50<\/td>\n<td>2.11<\/td>\n<td>4.70<\/td>\n<td>6.79<\/td>\n<td>4.63<\/td>\n<td>6.68<\/td>\n<td>10.31<\/td>\n<\/tr>\n<tr>\n<td><strong><code>Rumble-2-Flash-AD-Q5_K_M.gguf<\/code><\/strong><\/td>\n<td><strong>0.28<\/strong><\/td>\n<td>6.94<\/td>\n<td>2.09<\/td>\n<td>4.78<\/td>\n<td>6.83<\/td>\n<td>4.51<\/td>\n<td>6.60<\/td>\n<td>10.52<\/td>\n<\/tr>\n<tr>\n<td><em>stock <code>Q6_K<\/code> (baseline, not shipped)<\/em><\/td>\n<td><em>0.27<\/em><\/td>\n<td><em>6.56<\/em><\/td>\n<td><strong><em>2.20<\/em><\/strong><\/td>\n<td><em>4.64<\/em><\/td>\n<td><em>6.84<\/em><\/td>\n<td><em>4.51<\/em><\/td>\n<td><em>6.66<\/em><\/td>\n<td><em>10.44<\/em><\/td>\n<\/tr>\n<tr>\n<td><em>our <code>AD-Q4_K_M<\/code> (tried, not shipped)<\/em><\/td>\n<td><em>0.26<\/em><\/td>\n<td><em>6.48<\/em><\/td>\n<td><strong><em>2.22<\/em><\/strong><\/td>\n<td><em>4.81<\/em><\/td>\n<td><em>7.05<\/em><\/td>\n<td><em>4.74<\/em><\/td>\n<td><strong><em>6.88<\/em><\/strong><\/td>\n<td><em>10.54<\/em><\/td>\n<\/tr>\n<tr>\n<td><em>stock <code>Q5_K_M<\/code> (baseline, not shipped)<\/em><\/td>\n<td><em>0.24<\/em><\/td>\n<td>5.92<\/td>\n<td>2.14<\/td>\n<td>4.89<\/td>\n<td>7.00<\/td>\n<td>4.57<\/td>\n<td><strong><em>6.87<\/em><\/strong><\/td>\n<td><strong><em>10.76<\/em><\/strong><\/td>\n<\/tr>\n<tr>\n<td><em>stock <code>Q4_K_M<\/code> (baseline, not shipped)<\/em><\/td>\n<td><em>0.22<\/em><\/td>\n<td>5.32<\/td>\n<td><strong><em>2.42<\/em><\/strong><\/td>\n<td><strong><em>5.21<\/em><\/strong><\/td>\n<td><strong><em>7.62<\/em><\/strong><\/td>\n<td><strong><em>5.14<\/em><\/strong><\/td>\n<td><strong><em>7.47<\/em><\/strong><\/td>\n<td><strong><em>11.99<\/em><\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>As a result of the comparison, it can be seen that the recommended <code>AD-Q5_K_M<\/code> maintains performance within or extremely close to <code>BF16<\/code> (original precision) across most test sets. On the other hand, quantizations aimed at higher compression rates, such as <code>Q4_K_M<\/code>, tend to show a decrease in accuracy across all test sets. In particular, for Mandarin CER, a degradation of up to 1.6 points is reported with <code>Q4_K_M<\/code>.<\/p>\n<p>Additionally, it has been reported that the standard implementation (stock) of <code>llama.cpp<\/code> passes zero-padding at the end of audio clips to the model as sound, causing shorter clips to transcribe words that were not actually spoken. The figures in this documentation include a fix applied to resolve this issue. With this fix, <code>llama.cpp<\/code> is reported to exhibit performance that matches the Transformers reference model within the margin of noise.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>This model specializes in Automatic Speech Recognition (ASR), converting audio data into text. It supports as many as 45 languages and is intended for use in a wide range of linguistic environments. Specifically, it is suited for the following use cases:<\/p>\n<ul>\n<li>Multilingual audio transcription<\/li>\n<li>Converting voice data into text<\/li>\n<li>Local offline speech recognition<\/li>\n<\/ul>\n<p>Furthermore, it is built to run on <code>llama.cpp<\/code>, allowing for detailed transcription controls via <code>chat_template_kwargs<\/code>, such as a <code>digits<\/code> setting to output numbers in numeric format, retaining or removing fillers like &#8220;um&#8221; and &#8220;ah&#8221;, and condensing repeated words.<\/p>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 509M parameters (taken from the base model Reson8\/Rumble-2-Flash)<\/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>0.6GB<\/td>\n<td>0.7GB<\/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>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:peers --><\/p>\n<h2>Recent Models in the Same Size Class<\/h2>\n<p><em>Models with <\/em><em>up to 4B<\/em><em> parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site&#8217;s estimates; licenses are as stated on the model cards.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Parameters<\/th>\n<th>Smallest VRAM tier<\/th>\n<th>License<\/th>\n<th>Our article<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>ibm-granite\/granite-speech-5.0-2b-ONNX<\/td>\n<td>2.3B<\/td>\n<td>8GB<\/td>\n<td>apache-2.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/ibm-granite-speech-5-0-2b-onnx\/\">granite-speech-5.0-2b-ONNX Speech Recognition Model: 8GB+ VRAM<\/a> (2026-11-17)<\/td>\n<\/tr>\n<tr>\n<td>ggml-org\/higgs-audio-v3-stt-GGUF<\/td>\n<td>2.7B<\/td>\n<td>4GB<\/td>\n<td>apache-2.0<\/td>\n<td><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> (2026-11-17)<\/td>\n<\/tr>\n<tr>\n<td>FermionResearch\/Phonon-2<\/td>\n<td>627M<\/td>\n<td>4GB<\/td>\n<td>cc-by-4.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/01\/phonon-2-open-weight-english-asr-model\/\">Phonon-2 Speech Recognition Model: 4GB+ VRAM<\/a> (2026-09-30)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:peers --><\/p>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/AtomicChat\/Rumble-2-Flash-GGUF\/tree\/main\">AtomicChat\/Rumble-2-Flash-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>mmproj-Rumble-2-Flash-F16.gguf<\/code><\/td>\n<td>379MB<\/td>\n<\/tr>\n<tr>\n<td><code>mmproj-Rumble-2-Flash-Q8_0.gguf<\/code><\/td>\n<td>214MB<\/td>\n<\/tr>\n<tr>\n<td><code>Rumble-2-Flash-AD-Q5_K_M.gguf<\/code><\/td>\n<td>283MB<\/td>\n<\/tr>\n<tr>\n<td><code>Rumble-2-Flash-BF16.gguf<\/code><\/td>\n<td>648MB<\/td>\n<\/tr>\n<tr>\n<td><code>Rumble-2-Flash-Q8_0.gguf<\/code><\/td>\n<td>346MB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>This model is distributed in <code>GGUF<\/code> format. Using it requires two files: the main model file (<code>.gguf<\/code>) and the <code>mmproj<\/code> audio encoder file. The recommended combination is the pair of <code>Rumble-2-Flash-AD-Q5_K_M.gguf<\/code> and <code>mmproj-Rumble-2-Flash-Q8_0.gguf<\/code>.<\/p>\n<p>You can run it using <code>llama.cpp<\/code> with a command like the following:<\/p>\n<pre><code class=\"language-bash\">llama-server -m Rumble-2-Flash-AD-Q5_K_M.gguf --mmproj mmproj-Rumble-2-Flash-Q8_0.gguf -ngl 99\n<\/code><\/pre>\n<p>When using the command-line interface (<code>llama-mtmd-cli<\/code>), it is recommended to run it with the <code>--jinja<\/code> flag to correctly apply the embedded template.<\/p>\n<pre><code class=\"language-bash\">llama-mtmd-cli -m Rumble-2-Flash-AD-Q5_K_M.gguf --mmproj mmproj-Rumble-2-Flash-Q8_0.gguf --jinja --audio audio.wav -p &quot;lang:en&quot; --temp 0\n<\/code><\/pre>\n<p>The repository is available on Hugging Face at <code>AtomicChat\/Rumble-2-Flash-GGUF<\/code>.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/19\/aion-1-0-instruct-2-3b-gguf-released\/\">Aion-1.0-Instruct-2.3B-GGUF 2.3B Text Generation Model: 4GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/18\/maxwell-v1-flash-gguf-released-specs-performance-and-setup\/\">Maxwell-V1-Flash-GGUF 35B MoE Model: 24GB+ 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>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<li><strong>How to read WER, CER<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/AtomicChat\/Rumble-2-Flash-GGUF\">https:\/\/huggingface.co\/AtomicChat\/Rumble-2-Flash-GGUF<\/a><\/li>\n<\/ul>\n<blockquote>\n<p><strong>This article contains unverified information.<\/strong> We will append an update note once it is confirmed by a primary source.<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Explore the GGUF release of Rumble-2-Flash by AtomicChat, supporting 45 languages for local offline ASR with llama.cpp.<\/p>\n","protected":false},"author":1,"featured_media":20061,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[5297,163,1073,5380,1565,1017],"class_list":["post-20062","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-atomicchat-en","tag-gguf-en","tag-llama-cpp-en","tag-rumble-2-flash-en","tag-unverified","tag--en"],"lang":"en","translations":{"en":20062,"ja":20060},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/20062","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=20062"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/20062\/revisions"}],"predecessor-version":[{"id":20270,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/20062\/revisions\/20270"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/20061"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=20062"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=20062"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=20062"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}