{"id":21659,"date":"2026-11-21T06:11:55","date_gmt":"2026-11-20T21:11:55","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/21\/tml-d24-varloop-research-gguf-2\/"},"modified":"2026-11-22T06:47:28","modified_gmt":"2026-11-21T21:47:28","slug":"tml-d24-varloop-research-gguf-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/21\/tml-d24-varloop-research-gguf-2\/","title":{"rendered":"TML-d24-varloop-research-GGUF Vision-Language Model: 24GB+ 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\/ggml-org\/TML-d24-varloop-research-GGUF\">ggml-org\/TML-d24-varloop-research-GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-20<\/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<p><!-- lmw:lab-summary --><\/p>\n<p><strong>What we checked ourselves<\/strong><\/p>\n<ul>\n<li>It needs about 17% more tokens than the Qwen3 tokenizer for the same Japanese text, so only about 0.85\u00d7 as much Japanese fits in the same context length.<\/li>\n<\/ul>\n<p>Details and conditions are in \u201cOur Own Measurements\u201d below.<\/p>\n<p><!-- \/lmw:lab-summary --><\/p>\n<h2>Overview<\/h2>\n<p>ggml-org has released &#8220;ggml-org\/TML-d24-varloop-research-GGUF&#8221;, which is a GGUF quantized version of the research multimodal model developed by Thinking Machines Lab. This model is based on the original model <a href=\"https:\/\/huggingface.co\/thinkingmachines\/TML-d24-varloop-research\">thinkingmachines\/TML-d24-varloop-research<\/a> and supports both text and image inputs. It adopts an architecture that allows changing the number of loop layer passes during inference, and is provided as a research model to verify how additional computational load affects internal processing and capabilities.<\/p>\n<h2>Specifications<\/h2>\n<ul>\n<li>Parameters: Total parameters 26.4B, active parameters 3.4B (at 1 loop pass; each additional loop pass increases effective active parameters by approximately 1.0B per token)<\/li>\n<li>Architecture: 24-layer Mixture-of-Experts (MoE). 8 out of 128 total experts are routed per token. Layers 6 to 17 operate as a variable loop core<\/li>\n<li>Context Length: Up to 1M tokens<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>Performance data for this model is published in the model card of the original model, &#8220;thinkingmachines\/TML-d24-varloop-research&#8221;. The table below shows the benchmark scores for each loop count (1 to 5 passes) in the original model and the measurement results of the comparison target &#8220;Inkling 2 Mini (3 passes)&#8221; (some columns from the original table have been omitted for readability).<\/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><\/th>\n<th>Inkling 2 Mini \/ (3 passes)<\/th>\n<th>TML-d24-varloop-research \/ 1 pass<\/th>\n<th>TML-d24-varloop-research \/ 3 passes<\/th>\n<th>TML-d24-varloop-research \/ 5 passes<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>Agentic Coding<\/strong><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>DeepSWE v1.1 mini-swe-agent v2.4.6<\/td>\n<td>48.2%<\/td>\n<td>0.7%<\/td>\n<td>47.4%<\/td>\n<td>48.2%<\/td>\n<\/tr>\n<tr>\n<td>SWEBench Pro v2 (Hard) mini-swe-agent v2.4.6<\/td>\n<td>77.9%<\/td>\n<td>39.7%<\/td>\n<td>78.9%<\/td>\n<td>78.4%<\/td>\n<\/tr>\n<tr>\n<td>SWEBench Multilingual bash only<\/td>\n<td>73.7%<\/td>\n<td>40.8%<\/td>\n<td>71.8%<\/td>\n<td>72.0%<\/td>\n<\/tr>\n<tr>\n<td>SWEBench Verified bash only<\/td>\n<td>79.2%<\/td>\n<td>56.3%<\/td>\n<td>78.7%<\/td>\n<td>79.6%<\/td>\n<\/tr>\n<tr>\n<td><strong>Agentic Search, Tool Use, and Knowledge Work<\/strong><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>BrowseComp original, no context management<\/td>\n<td>76.0%<\/td>\n<td>29.4%<\/td>\n<td>65.7%<\/td>\n<td>71.0%<\/td>\n<\/tr>\n<tr>\n<td>MCP Atlas Public<\/td>\n<td>82.7%<\/td>\n<td>70.0%<\/td>\n<td>79.4%<\/td>\n<td>80.0%<\/td>\n<\/tr>\n<tr>\n<td>JobBench<\/td>\n<td>49.6%<\/td>\n<td>22.8%<\/td>\n<td>63.8%<\/td>\n<td>63.3%<\/td>\n<\/tr>\n<tr>\n<td>EnterpriseOps-Gym LangChain ReAct loop<\/td>\n<td>36.7%<\/td>\n<td>25.7%<\/td>\n<td>36.6%<\/td>\n<td>37.6%<\/td>\n<\/tr>\n<tr>\n<td><strong>Vision Interpretation and Creation<\/strong><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>GDP.pdf<\/td>\n<td>16.8%<\/td>\n<td>4.6%<\/td>\n<td>19.6%<\/td>\n<td>20.0%<\/td>\n<\/tr>\n<tr>\n<td>PerceptionBench<\/td>\n<td>57.5%<\/td>\n<td>45.5%<\/td>\n<td>53.9%<\/td>\n<td>52.7%<\/td>\n<\/tr>\n<tr>\n<td>MMMU Pro Standard 10<\/td>\n<td>74.9%<\/td>\n<td>63.6%<\/td>\n<td>71.4%<\/td>\n<td>73.2%<\/td>\n<\/tr>\n<tr>\n<td>OmniDocBench v1.6<\/td>\n<td>92.8%<\/td>\n<td>86.6%<\/td>\n<td>92.1%<\/td>\n<td>92.5%<\/td>\n<\/tr>\n<tr>\n<td>BabyVision original<\/td>\n<td>53.4%<\/td>\n<td>32.7%<\/td>\n<td>50.3%<\/td>\n<td>49.2%<\/td>\n<\/tr>\n<tr>\n<td>BabyVision with Python<\/td>\n<td>88.8%<\/td>\n<td>67.2%<\/td>\n<td>88.9%<\/td>\n<td>89.5%<\/td>\n<\/tr>\n<tr>\n<td>ScreenSpot Pro original<\/td>\n<td>76.5%<\/td>\n<td>54.4%<\/td>\n<td>72.0%<\/td>\n<td>73.1%<\/td>\n<\/tr>\n<tr>\n<td>ScreenSpot Pro with Python<\/td>\n<td>87.7%<\/td>\n<td>80.7%<\/td>\n<td>87.7%<\/td>\n<td>88.1%<\/td>\n<\/tr>\n<tr>\n<td>Charxiv RQ original<\/td>\n<td>80.6%<\/td>\n<td>75.0%<\/td>\n<td>79.9%<\/td>\n<td>80.4%<\/td>\n<\/tr>\n<tr>\n<td>Charxiv RQ with Python<\/td>\n<td>88.3%<\/td>\n<td>84.7%<\/td>\n<td>87.6%<\/td>\n<td>87.9%<\/td>\n<\/tr>\n<tr>\n<td>Chartography original<\/td>\n<td>22.1%<\/td>\n<td>14.5%<\/td>\n<td>21.6%<\/td>\n<td>22.6%<\/td>\n<\/tr>\n<tr>\n<td>Chartography with Python<\/td>\n<td>72.5%<\/td>\n<td>49.1%<\/td>\n<td>72.3%<\/td>\n<td>72.3%<\/td>\n<\/tr>\n<tr>\n<td>ZeroBench (main) original<\/td>\n<td>17.0%<\/td>\n<td>9.0%<\/td>\n<td>19.0%<\/td>\n<td>17.0%<\/td>\n<\/tr>\n<tr>\n<td>ZeroBench (main) with Python<\/td>\n<td>34.0%<\/td>\n<td>16.0%<\/td>\n<td>40.0%<\/td>\n<td>37.0%<\/td>\n<\/tr>\n<tr>\n<td>Humanity\u2019s Sixth Sense<\/td>\n<td>23.8%<\/td>\n<td>22.4%<\/td>\n<td>25.5%<\/td>\n<td>27.0%<\/td>\n<\/tr>\n<tr>\n<td>MathVision<\/td>\n<td>89.0%<\/td>\n<td>74.8%<\/td>\n<td>86.6%<\/td>\n<td>88.1%<\/td>\n<\/tr>\n<tr>\n<td>MuirBench<\/td>\n<td>77.5%<\/td>\n<td>73.6%<\/td>\n<td>77.1%<\/td>\n<td>77.8%<\/td>\n<\/tr>\n<tr>\n<td>BVB<\/td>\n<td>68.4%<\/td>\n<td>53.9%<\/td>\n<td>68.5%<\/td>\n<td>69.5%<\/td>\n<\/tr>\n<tr>\n<td><strong>Reasoning, Knowledge, and Epistemics<\/strong><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>HLE-Diamond<\/td>\n<td>15.2%<\/td>\n<td>6.3%<\/td>\n<td>9.5%<\/td>\n<td>11.1%<\/td>\n<\/tr>\n<tr>\n<td>HLE (text only)<\/td>\n<td>30.7%<\/td>\n<td>10.5%<\/td>\n<td>24.9%<\/td>\n<td>25.9%<\/td>\n<\/tr>\n<tr>\n<td>GPQA Diamond<\/td>\n<td>86.6%<\/td>\n<td>69.6%<\/td>\n<td>84.2%<\/td>\n<td>83.1%<\/td>\n<\/tr>\n<tr>\n<td>Apex 2025<\/td>\n<td>34.4%<\/td>\n<td>3.1%<\/td>\n<td>19.8%<\/td>\n<td>18.8%<\/td>\n<\/tr>\n<tr>\n<td>ComplexConstraints v1.1 openhands v1.45.0<\/td>\n<td>33.4%<\/td>\n<td>4.5%<\/td>\n<td>37.9%<\/td>\n<td>38.9%<\/td>\n<\/tr>\n<tr>\n<td>SimpleQA Verified<\/td>\n<td>8.4%<\/td>\n<td>6.3%<\/td>\n<td>7.3%<\/td>\n<td>6.7%<\/td>\n<\/tr>\n<tr>\n<td>AA-Omniscience v1.1 (Public) accuracy<\/td>\n<td>21.3%<\/td>\n<td>15.5%<\/td>\n<td>18.4%<\/td>\n<td>17.4%<\/td>\n<\/tr>\n<tr>\n<td>AA-Omniscience v1.1 (Public) non-hallucination rate<\/td>\n<td>64.6%<\/td>\n<td>79.3%<\/td>\n<td>84.0%<\/td>\n<td>85.9%<\/td>\n<\/tr>\n<tr>\n<td>SimpleVQA<\/td>\n<td>45.4%<\/td>\n<td>39.0%<\/td>\n<td>42.4%<\/td>\n<td>41.3%<\/td>\n<\/tr>\n<tr>\n<td>WorldVQA<\/td>\n<td>15.2%<\/td>\n<td>11.3%<\/td>\n<td>14.2%<\/td>\n<td>13.5%<\/td>\n<\/tr>\n<tr>\n<td>GraphWalks BFS 256K<\/td>\n<td>69.7%<\/td>\n<td>23.5%<\/td>\n<td>64.3%<\/td>\n<td>63.8%<\/td>\n<\/tr>\n<tr>\n<td>MRCR v2 8-needle 256K\u2013512K<\/td>\n<td>46.4%<\/td>\n<td>27.2%<\/td>\n<td>89.2%<\/td>\n<td>92.1%<\/td>\n<\/tr>\n<tr>\n<td><strong>Safety and Alignment<\/strong><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>FORTRESS (adversarial) Public<\/td>\n<td>89.5%<\/td>\n<td>89.0%<\/td>\n<td>90.0%<\/td>\n<td>89.4%<\/td>\n<\/tr>\n<tr>\n<td>FORTRESS (benign) Public<\/td>\n<td>96.4%<\/td>\n<td>95.0%<\/td>\n<td>95.3%<\/td>\n<td>95.0%<\/td>\n<\/tr>\n<tr>\n<td>StrongREJECT (none)<\/td>\n<td>97.8%<\/td>\n<td>98.8%<\/td>\n<td>98.2%<\/td>\n<td>98.4%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to the publisher&#8217;s measurement results, the model exhibits a characteristic where increasing the number of loop passes significantly improves scores in complex domains such as coding and reasoning. On &#8220;SWEBench Verified bash only&#8221;, which measures the ability to fix real GitHub issues, accuracy rises from 56.3% with 1 pass to 78.7% with 3 passes and 79.6% with 5 passes, matching the high accuracy of Inkling 2 Mini (79.2%). Additionally, on &#8220;GPQA Diamond&#8221;, which handles advanced PhD-level problems, the Pass@1 accuracy jumps from 69.6% with 1 pass to 84.2% with 3 passes. On the other hand, safety and alignment metrics such as &#8220;FORTRESS (adversarial) Public&#8221; show no significant changes, recording 89.0% for 1 pass, 90.0% for 3 passes, and 89.4% for 5 passes. Thus, while increasing loop computations drastically enhances performance on complex thinking tasks, behavior in areas like safety maintains stable scores.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>This model is a conversational multimodal model that accepts both text and image inputs. It is provided as a research model to evaluate changes in inference compute and internal model behavior within a variable loop structure. Its strengths and intended use cases are based on the characteristics of the original model &#8220;thinkingmachines\/TML-d24-varloop-research&#8221;.<\/p>\n<h3>Adjusting Inference Compute According to Tasks<\/h3>\n<p>Because the number of loop layer passes can be modified during inference, computational load can be adjusted to match task difficulty. For coding tasks involving automatic code correction and repository operations, and reasoning tasks requiring complex logical flow, increasing the loop count can extract higher processing accuracy.<\/p>\n<h3>Integrated Understanding and Dialogue of Images and Text<\/h3>\n<p>It supports pixel-based image input (recommended resolution: 40px to 4096px per side). Visual information inputted via the Vision Transformer is projected into a shared latent space along with text and processed in an integrated manner. It can be used for dialogues involving visual elements such as document image reading, chart analysis, and screen state comprehension.<\/p>\n<h3>Parameter Control of Reasoning Depth (<code>reasoning_effort<\/code>)<\/h3>\n<p>It supports the <code>reasoning_effort<\/code> parameter for setting the model&#8217;s thinking depth. You can specify string labels such as <code>none<\/code>, <code>minimal<\/code>, <code>low<\/code>, <code>medium<\/code>, <code>high<\/code> (default), <code>xhigh<\/code>, and <code>max<\/code>, or a numerical value from 0.0 to 0.99 to adjust inference depth.<\/p>\n<p>Note that this model is positioned for research purposes and has not been tuned for production deployment. For general practical use, utilizing production-grade models (such as Inkling 2 Mini) is recommended, and this model is primarily designed for investigating and experimenting with loop computation structures during inference.<\/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>24GB (RTX 4090 \/ 3090, etc.)<\/td>\n<td>IQ4_NL<\/td>\n<td>16.4GB<\/td>\n<td>19.7GB<\/td>\n<\/tr>\n<tr>\n<td>32GB (RTX 5090, etc.)<\/td>\n<td>Q8_0<\/td>\n<td>25.9GB<\/td>\n<td>31.1GB<\/td>\n<\/tr>\n<tr>\n<td>80GB class (A100 \/ H100)<\/td>\n<td>BF16<\/td>\n<td>48.6GB<\/td>\n<td>58.4GB<\/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><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.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:lab --><\/p>\n<h2>Our Own Measurements<\/h2>\n<p>Values we checked ourselves on our server (no GPU) by actually reading this model&#8217;s files, without running the model \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>Measurements We Did Not Take<\/h3>\n<p>This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation). Only values we checked without running the model, such as token counts and the GGUF header, are shown.<\/p>\n<h3>Japanese Token Efficiency<\/h3>\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>Tokenizer<\/th>\n<th>Tokens per 1,000 Japanese characters<\/th>\n<th>Ratio to the same text in English<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>This model<\/strong><\/td>\n<td><strong>806<\/strong><\/td>\n<td><strong>1.38\u00d7<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Qwen3<\/td>\n<td>688<\/td>\n<td>1.26\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Llama 3.2<\/td>\n<td>744<\/td>\n<td>1.36\u00d7<\/td>\n<\/tr>\n<tr>\n<td>Gemma 3<\/td>\n<td>564<\/td>\n<td>1.03\u00d7<\/td>\n<\/tr>\n<tr>\n<td>gpt-oss<\/td>\n<td>795<\/td>\n<td>1.45\u00d7<\/td>\n<\/tr>\n<tr>\n<td>LLM-jp-3<\/td>\n<td>497<\/td>\n<td>0.85\u00d7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>It needs about 17% more tokens than the Qwen3 tokenizer for the same Japanese text, so only about 0.85\u00d7 as much Japanese fits in the same context length.<\/p>\n<p>Counted with the <code>tokenizer.json<\/code> of <a href=\"https:\/\/huggingface.co\/thinkingmachines\/TML-d24-varloop-research\">thinkingmachines\/TML-d24-varloop-research<\/a> on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.<\/p>\n<h3>Inside the GGUF File<\/h3>\n<p>File: <a href=\"https:\/\/huggingface.co\/ggml-org\/TML-d24-varloop-research-GGUF\/blob\/main\/TML-d24-varloop-research-IQ4_NL.gguf\">TML-d24-varloop-research-IQ4_NL.gguf<\/a> (16.42GB, <code>IQ4_NL<\/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>inkling2_loop<\/code><\/td>\n<\/tr>\n<tr>\n<td>Maximum trained context length<\/td>\n<td>1,048,576 tokens<\/td>\n<\/tr>\n<tr>\n<td>Layers<\/td>\n<td>24<\/td>\n<\/tr>\n<tr>\n<td>Experts<\/td>\n<td>8 active out of 128<\/td>\n<\/tr>\n<tr>\n<td>Vocabulary size<\/td>\n<td>200,058<\/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>Used (<code>inkling2-loop-calibration-v6.txt<\/code>, 1,770 chunks)<\/td>\n<\/tr>\n<tr>\n<td>Weight types (share of parameters)<\/td>\n<td>IQ4_NL 78.2% \/ Q8_0 21.7% \/ F32 0.1%<\/td>\n<\/tr>\n<tr>\n<td>Average bits per weight<\/td>\n<td>5.41 bits<\/td>\n<\/tr>\n<tr>\n<td>Embedding \/ output layer type<\/td>\n<td>Q8_0 \/ Q8_0<\/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<p><!-- \/lmw:lab --><\/p>\n<h2>How to Get It<\/h2>\n<p>This model, &#8220;ggml-org\/TML-d24-varloop-research-GGUF&#8221;, is available in GGUF format on the Hugging Face repository. Gated license agreement procedures for downloading are not required.<\/p>\n<p>Compatible engines such as <code>llama.cpp<\/code> or <code>llama.app<\/code> can be used as execution environments. To launch an inference server, run the following command:<\/p>\n<pre><code class=\"language-bash\">llama serve -hf ggml-org\/TML-d24-varloop-research-GGUF:IQ4_NL\n<\/code><\/pre>\n<p>If you want to change the number of loop layer passes to an arbitrary value (e.g., 2 passes) when starting the server, add the <code>--override-kv<\/code> option as follows:<\/p>\n<pre><code class=\"language-bash\">llama serve -hf ggml-org\/TML-d24-varloop-research-GGUF:IQ4_NL --override-kv inkling2_loop.loop.count=int:2\n<\/code><\/pre>\n<p>The distribution formats include <code>IQ4_NL<\/code> calibrated with imatrix, as well as <code>Q8_0<\/code> and <code>BF16<\/code>. In the IQ4_NL quantized version, the expert weights of layers 0 to 19 are quantized to IQ4_NL, while the F32 weights of the loop controller and other weight matrices are maintained at Q8_0 precision.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/qwen3-coder-30b-a3b-instruct-gguf-2\/\">Qwen3-Coder-30B-A3B-Instruct-GGUF Text Generation Model: 24GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/vllm-d3-gguf-models-released\/\">vllm-d3 GGUF Models Released: Decision Models for Local Run<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/02\/ggml-org-openjev-gguf-released-2\/\">OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM<\/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 24GB tier) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-24gb-en\/\">Other models that run on a 24GB GPU<\/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 IQ4_NL, 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<li><strong>How to read SWE-bench, MMMU, HLE<\/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\/ggml-org\/TML-d24-varloop-research-GGUF\">https:\/\/huggingface.co\/ggml-org\/TML-d24-varloop-research-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/thinkingmachines\/TML-d24-varloop-research\">https:\/\/huggingface.co\/thinkingmachines\/TML-d24-varloop-research<\/a><\/li>\n<\/ul>\n<p><!-- lmw:updates --><\/p>\n<h2>Update History<\/h2>\n<ul>\n<li>2026-11-22: Added our own measurements: Japanese token efficiency, what is inside the GGUF file.<\/li>\n<\/ul>\n<p><!-- \/lmw:updates --><\/p>\n","protected":false},"excerpt":{"rendered":"<p>GGUF quantization of Thinking Machines Lab&#8217;s research multimodal model TML-d24-varloop-research is now available via ggml-org.<\/p>\n","protected":false},"author":1,"featured_media":21658,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[161,1970,163,1073,165,5519,1547,1952],"class_list":["post-21659","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-ggml-en","tag-ggml-org-en","tag-gguf-en","tag-llama-cpp-en","tag-moe-en","tag-tml-d24-varloop-research-en","tag-verified","tag-vlm-en"],"lang":"en","translations":{"en":21659,"ja":21657},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21659","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=21659"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21659\/revisions"}],"predecessor-version":[{"id":22207,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21659\/revisions\/22207"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/21658"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=21659"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=21659"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=21659"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}