output.md

{
    "title": "Qwen2.5-VL-32B: Smarter and Lighter",
    "byline": "Simon Willison",
    "dir": null,
    "lang": "en-gb",
    "content": "<div id=\"readability-page-1\" class=\"page\"><div>\n\n\n\n<p><strong><a href=\"https://qwenlm.github.io/blog/qwen2.5-vl-32b/\">Qwen2.5-VL-32B: Smarter and Lighter</a></strong>. The second big open weight LLM release from China today - the first being <a href=\"https://simonwillison.net/2025/Mar/24/deepseek/\">DeepSeek v3-0324</a>.</p>\n<p>Qwen's previous vision model was Qwen2.5 VL, <a href=\"https://simonwillison.net/2025/Jan/27/qwen25-vl-qwen25-vl-qwen25-vl/\">released in January</a> in 3B, 7B and 72B sizes.</p>\n<p>Today's Apache 2.0 licensed release is a 32B model, which is quickly becoming my personal favourite model size - large enough to have GPT-4-class capabilities, but small enough that on my 64GB Mack there's still enough RAM for me to run other memory-hungry applications like Firefox and VS Code.</p>\n<p>Qwen claim that the new model (when compared to their previous 2.5 VL family) can \"align more closely with human preferences\", is better at \"mathematical reasoning\" and provides \"enhanced accuracy and detailed analysis in tasks such as image parsing, content recognition, and visual logic deduction\".</p>\n<p>They also offer some presumably carefully selected benchmark results showing it out-performing Gemma 3-27B, Mistral Small 3.1 24B and GPT-4o-0513 (there have been two more recent GPT-4o releases since that one, 2024-08-16 and 2024-11-20).</p>\n<p>As usual, Prince Canuma had MLX versions of the models live within hours of the release, in <a href=\"https://huggingface.co/mlx-community/Qwen2.5-VL-32B-Instruct-4bit\">4 bit</a>, <a href=\"https://huggingface.co/mlx-community/Qwen2.5-VL-32B-Instruct-6bit\">6 bit</a>, <a href=\"https://huggingface.co/mlx-community/Qwen2.5-VL-32B-Instruct-8bit\">8 bit</a>, and <a href=\"https://huggingface.co/mlx-community/Qwen2.5-VL-32B-Instruct-bf16\">bf16</a> variants.</p>\n<p>I ran the 4bit version (a 18GB model download) using <code>uv</code> and Prince's <a href=\"https://github.com/Blaizzy/mlx-vlm\">mlx-vlm</a> like this:</p>\n<div><pre>uv run --with <span><span>'</span>numpy&lt;2<span>'</span></span> --with mlx-vlm \\\n  python -m mlx_vlm.generate \\\n    --model mlx-community/Qwen2.5-VL-32B-Instruct-4bit \\\n    --max-tokens 1000 \\\n    --temperature 0.0 \\\n    --prompt <span><span>\"</span>Describe this image.<span>\"</span></span> \\\n    --image Mpaboundrycdfw-1.png</pre></div>\n\n<p>Here's the image:</p>\n<p><img alt=\"A map image, described below\" src=\"https://static.simonwillison.net/static/2025/Mpaboundrycdfw-1.png\"></p>\n<p>And here's the resulting description, which I found to be <em>very</em> impressive:</p>\n<blockquote>\n<p>This image appears to be a map highlighting specific areas along the coastline, likely in California, based on the geographical features and place names. Here is a detailed description of the image:</p>\n<ol>\n<li>\n<p><strong>Geographic Features</strong>:</p>\n<ul>\n<li>The map shows a section of the coastline, with land on the right and the sea on the left.</li>\n<li>The land area includes parts of SanMateo County, as noted in the top-right corner.</li>\n<li>The coastline features several named locations, including:\n<ul>\n<li><strong>Point Montara</strong></li>\n<li><strong>Pillar Point</strong></li>\n<li><strong>Kings Rock</strong></li>\n<li><strong>VicenteCreek</strong></li>\n<li><strong>DennistonCreek</strong></li>\n<li><strong>State Beach</strong></li>\n</ul>\n</li>\n</ul>\n</li>\n<li>\n<p><strong>Highlighted Areas</strong>:</p>\n<ul>\n<li>The map contains two distinct areas marked with different colors:\n<ul>\n<li>\n<strong>Red Area (Top Section)</strong>\n<ul>\n<li>Labeled as \"MONTARA\" and \"MONTARA M R\" (possibly referring to a \"Marine Reserves\" or similar protected area).</li>\n<li>This area is marked with a dark red color and extends from the coastline inland.</li>\n</ul>\n</li>\n<li>\n<strong>Blue Area (Bottom Section)</strong>\n<ul>\n<li>Labeled as \"PILLAR\" and \"PILLAR P O I N T\" along with \"SMCA\" (Special Management Area).</li>\n<li>This area is marked with a blue color and is located closer to the coastline, including the area around \"Pillar Point.\"</li>\n</ul>\n</li>\n</ul>\n</li>\n</ul>\n</li>\n<li>\n<p><strong>Depth Contours</strong>:</p>\n<ul>\n<li>The map includes depth markings in the sea, indicating the bathymetrical features:\n<ul>\n<li>\n<strong>20 fathom (fm) line</strong> is marked in the upper-left part of the sea.</li>\n<li>\n<strong>10 fathom (fm) line</strong> is marked closer to the coastline.</li>\n<li>\n<strong>30 fathom (fm) line</strong> is marked further out to sea.</li>\n</ul>\n</li>\n</ul>\n</li>\n<li>\n<p><strong>Other Features</strong>:</p>\n<ul>\n<li>\n<strong>State Beach</strong>: Marked near the top-right, indicating a protected recreational area.</li>\n<li>\n<strong>Kings Rock</strong>: A prominent feature near the coastline, likely a rocky outcropping.</li>\n<li>\n<strong>Creeks</strong>: The map shows several creeks, including <strong>VicenteCreek</strong> and <strong>DennistonCreek</strong>, which flow into the sea.</li>\n</ul>\n</li>\n<li>\n<p><strong>Protected Areas</strong>:</p>\n<ul>\n<li>The map highlights specific protected areas:\n<ul>\n<li>\n<strong>Marine Reserves</strong>:\n<ul>\n<li>\"MONTARA M R\" (Marine Reserves) in red.</li>\n</ul>\n</li>\n<li>\n<strong>Special Management Area (SMCA)</strong>\n<ul>\n<li>\"PILLAR P O I N T\" in blue, indicating a Special Management Area.</li>\n</ul>\n</li>\n</ul>\n</li>\n</ul>\n</li>\n<li>\n<p><strong>Grid and Coordinates</strong>:</p>\n<ul>\n<li>The map includes a grid with latitude and longitude markings:\n<ul>\n<li>Latitude ranges from approximately 37\u00b025'N to 37\u00b035'N.</li>\n<li>Longitude ranges from approximately 122\u00b022.5'W to 122\u00b035.5'W.</li>\n</ul>\n</li>\n</ul>\n</li>\n<li>\n<p><strong>Topography</strong>:</p>\n<ul>\n<li>The land area shows topographic features, including elevations and vegetation, with green areas indicating higher elevations or vegetated land.</li>\n</ul>\n</li>\n<li>\n<p><strong>Other Labels</strong>:</p>\n<ul>\n<li>\n<strong>\"SMR\"</strong>: Likely stands for \"State Managed Reserves.\"</li>\n<li>\n<strong>\"SMCA\"</strong>: Likely stands for \"Special Management Control Area.\"</li>\n</ul>\n</li>\n</ol>\n<p>In summary, this map highlights specific protected areas along the coastline, including a red \"Marine Reserves\" area and a blue \"Special Management Area\" near \"Pillar Point.\" The map also includes depth markings, geographical features, and place names, providing a detailed view of the region's natural and protected areas.</p>\n</blockquote>\n\n<p>It included the following runtime statistics:</p>\n<pre><code>Prompt: 1051 tokens, 111.985 tokens-per-sec\nGeneration: 760 tokens, 17.328 tokens-per-sec\nPeak memory: 21.110 GB\n</code></pre>\n\n\n\n</div></div>",
    "textContent": "\n\n\n\nQwen2.5-VL-32B: Smarter and Lighter. The second big open weight LLM release from China today - the first being DeepSeek v3-0324.\nQwen's previous vision model was Qwen2.5 VL, released in January in 3B, 7B and 72B sizes.\nToday's Apache 2.0 licensed release is a 32B model, which is quickly becoming my personal favourite model size - large enough to have GPT-4-class capabilities, but small enough that on my 64GB Mack there's still enough RAM for me to run other memory-hungry applications like Firefox and VS Code.\nQwen claim that the new model (when compared to their previous 2.5 VL family) can \"align more closely with human preferences\", is better at \"mathematical reasoning\" and provides \"enhanced accuracy and detailed analysis in tasks such as image parsing, content recognition, and visual logic deduction\".\nThey also offer some presumably carefully selected benchmark results showing it out-performing Gemma 3-27B, Mistral Small 3.1 24B and GPT-4o-0513 (there have been two more recent GPT-4o releases since that one, 2024-08-16 and 2024-11-20).\nAs usual, Prince Canuma had MLX versions of the models live within hours of the release, in 4 bit, 6 bit, 8 bit, and bf16 variants.\nI ran the 4bit version (a 18GB model download) using uv and Prince's mlx-vlm like this:\nuv run --with 'numpy<2' --with mlx-vlm \\\n  python -m mlx_vlm.generate \\\n    --model mlx-community/Qwen2.5-VL-32B-Instruct-4bit \\\n    --max-tokens 1000 \\\n    --temperature 0.0 \\\n    --prompt \"Describe this image.\" \\\n    --image Mpaboundrycdfw-1.png\n\nHere's the image:\n\nAnd here's the resulting description, which I found to be very impressive:\n\nThis image appears to be a map highlighting specific areas along the coastline, likely in California, based on the geographical features and place names. Here is a detailed description of the image:\n\n\nGeographic Features:\n\nThe map shows a section of the coastline, with land on the right and the sea on the left.\nThe land area includes parts of SanMateo County, as noted in the top-right corner.\nThe coastline features several named locations, including:\n\nPoint Montara\nPillar Point\nKings Rock\nVicenteCreek\nDennistonCreek\nState Beach\n\n\n\n\n\nHighlighted Areas:\n\nThe map contains two distinct areas marked with different colors:\n\n\nRed Area (Top Section)\n\nLabeled as \"MONTARA\" and \"MONTARA M R\" (possibly referring to a \"Marine Reserves\" or similar protected area).\nThis area is marked with a dark red color and extends from the coastline inland.\n\n\n\nBlue Area (Bottom Section)\n\nLabeled as \"PILLAR\" and \"PILLAR P O I N T\" along with \"SMCA\" (Special Management Area).\nThis area is marked with a blue color and is located closer to the coastline, including the area around \"Pillar Point.\"\n\n\n\n\n\n\n\nDepth Contours:\n\nThe map includes depth markings in the sea, indicating the bathymetrical features:\n\n\n20 fathom (fm) line is marked in the upper-left part of the sea.\n\n10 fathom (fm) line is marked closer to the coastline.\n\n30 fathom (fm) line is marked further out to sea.\n\n\n\n\n\nOther Features:\n\n\nState Beach: Marked near the top-right, indicating a protected recreational area.\n\nKings Rock: A prominent feature near the coastline, likely a rocky outcropping.\n\nCreeks: The map shows several creeks, including VicenteCreek and DennistonCreek, which flow into the sea.\n\n\n\nProtected Areas:\n\nThe map highlights specific protected areas:\n\n\nMarine Reserves:\n\n\"MONTARA M R\" (Marine Reserves) in red.\n\n\n\nSpecial Management Area (SMCA)\n\n\"PILLAR P O I N T\" in blue, indicating a Special Management Area.\n\n\n\n\n\n\n\nGrid and Coordinates:\n\nThe map includes a grid with latitude and longitude markings:\n\nLatitude ranges from approximately 37\u00b025'N to 37\u00b035'N.\nLongitude ranges from approximately 122\u00b022.5'W to 122\u00b035.5'W.\n\n\n\n\n\nTopography:\n\nThe land area shows topographic features, including elevations and vegetation, with green areas indicating higher elevations or vegetated land.\n\n\n\nOther Labels:\n\n\n\"SMR\": Likely stands for \"State Managed Reserves.\"\n\n\"SMCA\": Likely stands for \"Special Management Control Area.\"\n\n\n\nIn summary, this map highlights specific protected areas along the coastline, including a red \"Marine Reserves\" area and a blue \"Special Management Area\" near \"Pillar Point.\" The map also includes depth markings, geographical features, and place names, providing a detailed view of the region's natural and protected areas.\n\n\nIt included the following runtime statistics:\nPrompt: 1051 tokens, 111.985 tokens-per-sec\nGeneration: 760 tokens, 17.328 tokens-per-sec\nPeak memory: 21.110 GB\n\n\n\n\n",
    "length": 4501,
    "excerpt": "The second big open weight LLM release from China today - the first being [DeepSeek v3-0324](https://simonwillison.net/2025/Mar/24/deepseek/). Qwen's previous vision model was Qwen2.5 VL, [released in January](https://simonwillison.net/2025/Jan/27/qwen25-vl-qwen25-vl-qwen25-vl/) in 3B, 7B and \u2026",
    "siteName": "Simon Willison\u2019s Weblog",
    "publishedTime": null
}