{"id":19402,"date":"2026-11-18T20:19:23","date_gmt":"2026-11-18T11:19:23","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/18\/sensenova-loop-9b-gguf-2\/"},"modified":"2026-11-19T06:51:41","modified_gmt":"2026-11-18T21:51:41","slug":"sensenova-loop-9b-gguf-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/18\/sensenova-loop-9b-gguf-2\/","title":{"rendered":"SenseNova-Loop-9B-GGUF Vision-Language Model: 8GB+ VRAM"},"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\/sensenova\/SenseNova-Loop-9B-GGUF\">sensenova\/SenseNova-Loop-9B-GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-sensenova-sensenova-loop-en\/\">SenseNova-Loop guide (2 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-18<\/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>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>SenseNova has released GGUF-format model weights for &#8220;SenseNova-Loop-9B&#8221;, a multimodal model equipped with a proprietary loop mechanism.<\/p>\n<p>This model is a multimodal model with approximately 9B parameters that accepts text and image inputs to generate text. Based on insights from the previously released Preview version, it adopts a looped architecture (Looped Transformer) that iteratively executes middle-layer weights and manifold-constrained hyper-connections (mHC), expanding computational depth within a limited parameter count.<\/p>\n<p>It is designed with agent tasks, coding, and data analysis tasks in mind, supporting a context length of up to 512K tokens. While maintaining a lightweight parameter scale, it aims for inference performance approaching larger models by iteratively processing the middle layers.<\/p>\n<h2>Specifications<\/h2>\n<p>The specifications listed in the model card are as follows:<\/p>\n<ul>\n<li>Parameter Count: Total parameters approx. 9B (inference-equivalent parameters approx. 14.6B)<\/li>\n<li>Architecture: Autoregressive multimodal decoder (loop block featuring 2-stream mHC)<\/li>\n<li>Input \/ Output: Text and images \/ Text<\/li>\n<li>Individual Decoder Layers: 32 layers<\/li>\n<li>Loop Target Layer Range \/ Passes: 24 layers in the [4, 28) range \/ 2 passes (weight sharing)<\/li>\n<li>Decoder Layer Executions: 56 layers per forward pass<\/li>\n<li>Attention Configuration: Hybrid configuration combining 3 linear attention layers and 1 full attention layer<\/li>\n<li>mHC Residual Streams: 2 streams (Sinkhorn normalization iterations: 20)<\/li>\n<li>Hidden Dimension: 4,096<\/li>\n<li>Feed-Forward Dimension: 12,288<\/li>\n<li>Vocabulary Size: 248,320<\/li>\n<li>Attention Heads: 16 full attention heads \/ 4 KV heads<\/li>\n<li>Weight Precision: Backbone BF16 \/ mHC routing parameters FP32<\/li>\n<li>Context Length: 512K tokens<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>The model card publishes comparison results for the official release version of &#8220;SenseNova-Loop-9B&#8221; against the earlier Preview version and other models.<\/p>\n<p>The benchmark measurement scores mentioned in the model card are as follows:<\/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 style=\"text-align: left;\">Benchmark<\/th>\n<th style=\"text-align: center;\">SenseNova-Loop-9B (Release)<\/th>\n<th style=\"text-align: center;\">SenseNova-Loop-9B-Preview<\/th>\n<th style=\"text-align: center;\">Qwen3.8-27B<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\">AutomationBench<\/td>\n<td style=\"text-align: center;\">36.67<\/td>\n<td style=\"text-align: center;\">14.33<\/td>\n<td style=\"text-align: center;\">36.67<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">Terminal-Bench 2.1<\/td>\n<td style=\"text-align: center;\">59.55<\/td>\n<td style=\"text-align: center;\">40.45<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">GAIA<\/td>\n<td style=\"text-align: center;\">69.90<\/td>\n<td style=\"text-align: center;\">37.86<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">AIDABench<\/td>\n<td style=\"text-align: center;\">65.31<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<td style=\"text-align: center;\">65.01<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">NL2Repo<\/td>\n<td style=\"text-align: center;\">41.36<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<td style=\"text-align: center;\">44.60<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">BIRD<\/td>\n<td style=\"text-align: center;\">72.34<\/td>\n<td style=\"text-align: center;\">71.38<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to measurements provided by the release source, the official version shows score improvements across all presented benchmarks compared to the Preview version. In particular, on GAIA\u2014which measures the ability to solve real-world research problems using tools such as search and file operations\u2014it shows a significant jump from 37.86 to 69.90. Additionally, on Terminal-Bench 2.1, which evaluates whether tasks can be completed by actually executing commands in a terminal, the score increases substantially from 40.45 to 59.55.<\/p>\n<p>In comparisons with Qwen3.8-27B, which has an approximately 3x larger parameter scale, it records an equivalent score of 36.67 on AutomationBench and slightly outperforms Qwen3.8-27B&#8217;s 65.01 with 65.31 on AIDABench. On the other hand, it stays at 41.36 on NL2Repo, falling short of Qwen3.8-27B&#8217;s 44.60. While it lags behind large models in specific repository generation tasks, it demonstrates scores competitive with 27B-class models in the domains of agent tasks and data analysis.<\/p>\n<p>Furthermore, the model card presents validation results with varying loop counts. It is reported that increasing the number of iterations from the standard 2 loops to 3\u20134 loops confirmed further score improvements in tasks such as LiveCodeBench V6 (which measures coding accuracy on new competitive programming problems) and AIDABench data visualization tasks.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>SenseNova-Loop-9B is a multimodal model that accepts both text and images as inputs and generates text. English and Chinese are tagged as supported languages, and it is expected to be applied to diverse tasks leveraging its extremely long context length of up to 512K tokens.<\/p>\n<p>The model card lists the following tasks as areas of strength for this model:<\/p>\n<ul>\n<li><strong>Agent Tasks and Tool Use<\/strong>: Suitable for complex workflows involving search, external tool calls, and terminal operations on the command line. It exhibits high performance in benchmarks for research and terminal operation tasks.<\/li>\n<li><strong>Data Analysis and Visualization<\/strong>: Supports analysis of tabular data and data visualization tasks, showing high aptitude in analytical evaluation metrics.<\/li>\n<li><strong>Coding and Design<\/strong>: Examples of visual and structural content creation are introduced, including code generation, SVG and website construction, 3D scene design, and poster layouts.<\/li>\n<li><strong>Long-Context Understanding<\/strong>: Supporting a context length of up to 512K tokens, it is suited for multi-step processing dealing with massive documents, codebases, and lengthy conversation histories.<\/li>\n<\/ul>\n<p>On the other hand, the release source notes the following issues as Known Limitations stemming from model scale and development timeframe:<\/p>\n<ul>\n<li><strong>Following Complex Instructions<\/strong>: When numerous complex tool definitions or multiple constraints are given simultaneously, it may misunderstand or overlook requirements.<\/li>\n<li><strong>Stability in Coding and Agent Tasks<\/strong>: Output consistency may be compromised in multi-step, long-term tasks or coding. While taking additional steps can sometimes allow it to self-correct errors, its capability in executing complex long-term tasks is reported to still have room for improvement.<\/li>\n<li><strong>Advanced Web Content Generation<\/strong>: When building rich websites containing complex 3D scenes or animations, it is difficult to complete everything in a single turn, requiring 2 to 3 iterations of corrections based on user feedback.<\/li>\n<\/ul>\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>8GB (RTX 4060 \/ 3060 Ti, etc.)<\/td>\n<td>Q5_K_M<\/td>\n<td>6.2GB<\/td>\n<td>7.5GB<\/td>\n<\/tr>\n<tr>\n<td>12GB (RTX 4070 \/ 3060 12GB, etc.)<\/td>\n<td>Q8_0<\/td>\n<td>9.1GB<\/td>\n<td>11.0GB<\/td>\n<\/tr>\n<tr>\n<td>24GB (RTX 4090 \/ 3090, etc.)<\/td>\n<td>BF16<\/td>\n<td>17.2GB<\/td>\n<td>20.6GB<\/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><strong>Runs in Ollama, LM Studio and llama.cpp as-is.<\/strong><\/p>\n<p>It is distributed in GGUF, so no conversion is needed.<\/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, converted builds we have found, and each engine&#8217;s own model registry. &#8220;Not found&#8221; means we have not seen such a build, not that none exists. 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:lab --><\/p>\n<h2>Our Own Measurements<\/h2>\n<p>Values we measured ourselves on our server (no GPU) by actually reading and running this model&#8217;s files \u2014 not figures copied from the model card. How we measure, and the results for every model: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/observations-en\/\">Our Measurements<\/a>.<\/p>\n<h3>What We Found<\/h3>\n<p>We attempted to run Sensenova-Loop-9B-GGUF with Q4_K_M quantization using only a CPU, but the current version of llama.cpp failed to load it because the model architecture was recognized as unknown. Although the model has a very large specification with a maximum context length of 524288 tokens during training, it ended in a state where execution is not even possible in our current environment. We need to wait for future updates on the llama.cpp side.<\/p>\n<h3>Inside the GGUF File<\/h3>\n<p>File: <a href=\"https:\/\/huggingface.co\/sensenova\/SenseNova-Loop-9B-GGUF\/blob\/main\/SenseNova-Loop-9B-Q4_K_M.gguf\">SenseNova-Loop-9B-Q4_K_M.gguf<\/a> (5.42GB, <code>Q4_K_M<\/code>). We read only the header (metadata) of the file, not the weights.<\/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>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Architecture (as named in the GGUF)<\/td>\n<td><code>sensenova_loop<\/code><\/td>\n<\/tr>\n<tr>\n<td>Maximum trained context length<\/td>\n<td>524,288 tokens<\/td>\n<\/tr>\n<tr>\n<td>Layers<\/td>\n<td>57<\/td>\n<\/tr>\n<tr>\n<td>Vocabulary size<\/td>\n<td>248,320<\/td>\n<\/tr>\n<tr>\n<td>Chat template<\/td>\n<td>Included (mentions tool calls, has a thinking switch)<\/td>\n<\/tr>\n<tr>\n<td>imatrix<\/td>\n<td>Not recorded in the file<\/td>\n<\/tr>\n<tr>\n<td>Weight types (share of parameters)<\/td>\n<td>Q4_K 73.9% \/ Q6_K 26.1% \/ F32 0.0%<\/td>\n<\/tr>\n<tr>\n<td>Average bits per weight<\/td>\n<td>5.05 bits<\/td>\n<\/tr>\n<tr>\n<td>Embedding \/ output layer type<\/td>\n<td>Q4_K \/ Q6_K<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.<\/p>\n<h3>Running It on a CPU Only<\/h3>\n<p>llama.cpp b12418 could not load this file: it does not know the architecture <code>sensenova_loop<\/code> yet. We will measure it again once support lands.<\/p>\n<p>Conditions: Neoverse-N1, 3 threads, no GPU, llama.cpp b12418, <code>Q4_K_M<\/code> (5.42GB) from <a href=\"https:\/\/huggingface.co\/sensenova\/SenseNova-Loop-9B-GGUF\">sensenova\/SenseNova-Loop-9B-GGUF<\/a>, measured 2026-11-19 (JST).<\/p>\n<p><!-- \/lmw:lab --><\/p>\n<h2>How to Get It<\/h2>\n<p>The GGUF version of SenseNova-Loop-9B is available in the Hugging Face <code>sensenova\/SenseNova-Loop-9B-GGUF<\/code> repository. The Apache-2.0 license is applied, and since it is not a gated model, anyone can directly download and use the weights without prior application or consent procedures.<\/p>\n<p>The following weights are provided in GGUF format:<\/p>\n<ul>\n<li>BF16<\/li>\n<li>Q8_0<\/li>\n<li>Q5_K_M<\/li>\n<li>Q4_K_M<\/li>\n<\/ul>\n<p>Note that because this model adopts a proprietary loop structure and mHC (manifold-constrained hyper-connections), it requires using a dedicated fork repository containing supporting code rather than standard llama.cpp. The release source advises building and using <a href=\"https:\/\/github.com\/OpenSenseNova\/llama.cpp\/tree\/24ea7bb78410e0f0f7bbb362e37c60cd9ecfc761\">OpenSenseNova\/llama.cpp<\/a>.<\/p>\n<p>An example command to load the Q4_K_M quantized weights and start the server after completing the build is as follows:<\/p>\n<pre><code class=\"language-bash\">llama serve -hf sensenova\/SenseNova-Loop-9B-GGUF:Q4_K_M \\\n  -ngl 99 -c 4096 --port 8000\n<\/code><\/pre>\n<p>Aside from the GGUF format, official distributions on Hugging Face include the base Safetensors BF16 weights (<code>sensenova\/SenseNova-Loop-9B<\/code>) and 4-bit AWQ format weights (<code>sensenova\/SenseNova-Loop-9B-AWQ<\/code>). Additionally, execution examples using Transformers (where <code>trust_remote_code=True<\/code> is mandatory to load custom code), LightLLM, and SGLang or vLLM via dedicated pull request branches are provided in the model card.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/18\/sensenova-loop-9b-awq-released\/\">SenseNova-Loop-9B-AWQ 512K Context Multimodal Model: 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 8GB 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>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-sensenova-sensenova-loop-en\/\">SenseNova-Loop family overview (2 articles, 1 converted builds)<\/a><\/li>\n<li><strong>Engines that run this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-llama-cpp-en\/\">llama.cpp<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-ollama-en\/\">Ollama<\/a><\/li>\n<li><strong>What Q5_K_M, 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-vision\">Other vision-language models<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/sensenova\/SenseNova-Loop-9B-GGUF\">sensenova\/SenseNova-Loop-9B-GGUF (Hugging Face)<\/a><\/li>\n<\/ul>\n<p><!-- lmw:updates --><\/p>\n<h2>Update History<\/h2>\n<ul>\n<li>2026-11-19: Added our own measurements: what is inside the GGUF file.<\/li>\n<\/ul>\n<p><!-- \/lmw:updates --><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Explore SenseNova-Loop-9B GGUF, a 9B multimodal model featuring a looped architecture and 512K context length for agent and coding tasks.<\/p>\n","protected":false},"author":1,"featured_media":19401,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[163,1073,5342,5344,5346,5348,1547,1952,896],"class_list":["post-19402","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-gguf-en","tag-llama-cpp-en","tag-mhc-en","tag-sensenova-en","tag-sensenova-loop-en","tag-sensenova-loop-9b-en","tag-verified","tag-vlm-en","tag--en"],"lang":"en","translations":{"en":19402,"ja":19400},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/19402","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=19402"}],"version-history":[{"count":5,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/19402\/revisions"}],"predecessor-version":[{"id":20129,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/19402\/revisions\/20129"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/19401"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=19402"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=19402"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=19402"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}