<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>gemma Archives - Urban Geo Analytics</title>
	<atom:link href="https://urbangeoanalytics.com/tag/gemma/feed/" rel="self" type="application/rss+xml" />
	<link></link>
	<description>Spatial Analysis, GeoAI &#38; Machine Learning</description>
	<lastBuildDate>Sat, 14 Nov 2026 08:10:56 +0000</lastBuildDate>
	<language>en-US</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=7.1.4</generator>

<image>
	<url>https://urbangeoanalytics.com/wp-content/uploads/2025/11/cropped-logo-urban-geo_512-32x32.png</url>
	<title>gemma Archives - Urban Geo Analytics</title>
	<link></link>
	<width>32</width>
	<height>32</height>
</image> 
	<item>
		<title>Do Open-Weight LLMs Reason From the Spatial Context They Are Given?</title>
		<link>https://urbangeoanalytics.com/llm-faithfulness-benchmark-geographic-reasoning/</link>
		
		<dc:creator><![CDATA[Joan Perez]]></dc:creator>
		<pubDate>Thu, 01 Oct 2026 06:19:39 +0000</pubDate>
				<category><![CDATA[Advanced]]></category>
		<category><![CDATA[Python]]></category>
		<category><![CDATA[Urbanism]]></category>
		<category><![CDATA[benchmark]]></category>
		<category><![CDATA[faithfulness]]></category>
		<category><![CDATA[gemma]]></category>
		<category><![CDATA[GeoAI]]></category>
		<category><![CDATA[GHS-POP]]></category>
		<category><![CDATA[hallucination]]></category>
		<category><![CDATA[Llama]]></category>
		<category><![CDATA[LLM]]></category>
		<category><![CDATA[network catchment]]></category>
		<category><![CDATA[open-weight models]]></category>
		<category><![CDATA[OpenStreetMap]]></category>
		<category><![CDATA[OSMnx]]></category>
		<category><![CDATA[Qwen3]]></category>
		<guid isPermaLink="false">https://urbangeoanalytics.com/?p=3118</guid>

					<description><![CDATA[<p>A new paper asks a question geographic evaluation has mostly skipped: once a language model is handed the right local data, does it reason from it, or fall back on what it already believes about the place? Sixteen open-weight configurations, three cities, ten seeds, and one planted false premise per case.</p>
<p>The post <a href="https://urbangeoanalytics.com/llm-faithfulness-benchmark-geographic-reasoning/">Do Open-Weight LLMs Reason From the Spatial Context They Are Given?</a> appeared first on <a href="https://urbangeoanalytics.com">Urban Geo Analytics</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class="fusion-fullwidth fullwidth-box fusion-builder-row-1 fusion-flex-container has-pattern-background has-mask-background nonhundred-percent-fullwidth non-hundred-percent-height-scrolling" style="--awb-border-radius-top-left:0px;--awb-border-radius-top-right:0px;--awb-border-radius-bottom-right:0px;--awb-border-radius-bottom-left:0px;--awb-flex-wrap:wrap;" id="contenu" ><div class="fusion-builder-row fusion-row fusion-flex-align-items-flex-start fusion-flex-content-wrap" style="max-width:1248px;margin-left: calc(-4% / 2 );margin-right: calc(-4% / 2 );"><div class="fusion-layout-column fusion_builder_column fusion-builder-column-0 fusion_builder_column_3_4 3_4 fusion-flex-column" style="--awb-bg-size:cover;--awb-width-large:75%;--awb-margin-top-large:0px;--awb-spacing-right-large:2.56%;--awb-margin-bottom-large:20px;--awb-spacing-left-large:2.56%;--awb-width-medium:75%;--awb-order-medium:0;--awb-spacing-right-medium:2.56%;--awb-spacing-left-medium:2.56%;--awb-width-small:100%;--awb-order-small:0;--awb-spacing-right-small:1.92%;--awb-spacing-left-small:1.92%;" id="contenu" data-scroll-devices="small-visibility,medium-visibility,large-visibility"><div class="fusion-column-wrapper fusion-column-has-shadow fusion-flex-justify-content-flex-start fusion-content-layout-column"><div class="fusion-text fusion-text-1 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>Language models know a lot about places, but they reason poorly over space and are unreliable when asked about a location from coordinates alone. The usual fix is to supply the relevant data in the prompt. This paper examines what happens next. Once the facts are in front of the model, does it use them, or does it override them with its own prior about the place?</p>
</div><div class="fusion-title title fusion-title-1 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">The paper</span></h2></div><div class="fusion-text fusion-text-2 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><div data-line="19" data-line-type="context" data-line-index="18">
<div data-line="10" data-line-type="context" data-line-index="9">&#8220;Network-based Spatial Context Retrieval for Open-weight LLMs: A Faithfulness Benchmark for Grounded Geographic Reasoning&#8221; separates two things that are usually measured together: being correct about the world, and being faithful to the context supplied.</div>
<div data-line="11" data-line-type="context" data-line-index="10"></div>
<div data-line="12" data-line-type="context" data-line-index="11">To do so, it fixes the context. For a point on a map, a short &#8220;spatial brief&#8221; is computed from OpenStreetMap and GHS-POP over the area reachable on foot: residents, density, buildings, street length, points of interest by category. The model receives finished numbers and only has to interpret them. Because the brief is known, every claim in an answer can be checked against it.</div>
<div data-line="13" data-line-type="context" data-line-index="12"></div>
<div data-line="14" data-line-type="context" data-line-index="13">The study applies this to three contrasting places, each tied to a theme in urban research that the models are likely to have met in training. The first is a neighbourhood on Chicago&#8217;s West Side, where 7,128 residents live within an 800 m walk that contains no supermarket or grocery store. The second is the historic core of Paris near Le Marais, with about 26,800 residents per km² and 971 points of interest within 400 m. The third is Hanoi&#8217;s Old Quarter, a district known for shops and tourism that nonetheless holds over 7,000 residents within a 300 m walk. The three are ordered by difficulty: Chicago asks the model to notice an absence, Paris to interpret a number, and Hanoi to hold a high resident count against a strong reputation.</div>
</div>
</div><div class="fusion-image-element awb-imageframe-style awb-imageframe-style-below awb-imageframe-style-1" style="text-align:center;--awb-margin-top:25px;--awb-margin-bottom:25px;--awb-caption-title-font-family:var(--body_typography-font-family);--awb-caption-title-font-weight:var(--body_typography-font-weight);--awb-caption-title-font-style:var(--body_typography-font-style);--awb-caption-title-size:var(--body_typography-font-size);--awb-caption-title-transform:var(--body_typography-text-transform);--awb-caption-title-line-height:var(--body_typography-line-height);--awb-caption-title-letter-spacing:var(--body_typography-letter-spacing);"><span class=" fusion-imageframe imageframe-none imageframe-1 hover-type-none"><img fetchpriority="high" decoding="async" width="1838" height="2000" alt="Figure 1: benchmark pipeline from walkable-area spatial brief to claim labelling and false-premise trap" title="fig1_pipeline" src="https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-scaled.png" class="img-responsive wp-image-3122" srcset="https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-200x218.png 200w, https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-400x435.png 400w, https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-600x653.png 600w, https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-800x871.png 800w, https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-1200x1306.png 1200w, https://urbangeoanalytics.com/wp-content/uploads/2026/10/fig1_pipeline-scaled.png 1838w" sizes="(max-width: 640px) 100vw, 1200px" /></span><div class="awb-imageframe-caption-container" style="text-align:center;"><div class="awb-imageframe-caption"><div class="awb-imageframe-caption-title">The two-stages pipeline</div></div></div></div><div class="fusion-title title fusion-title-2 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">How faithfulness is measured</span></h2></div><div class="fusion-text fusion-text-3 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>Answers are split into atomic claims. Each claim is labelled by its source, the brief or the model&#8217;s own knowledge, and by its correctness. This gives four categories: brief-true, brief-false, recall and hallucination.</p>
<p>Each conversation then ends with a trap. The user casually asserts something that sounds right but that the brief contradicts. In Chicago, that supermarkets are an easy walk away, in a catchment that has none. In Paris, that a district with 26,800 residents per km² is a calm, low-density corner. In Hanoi, that hardly anyone lives in the Old Quarter, where the brief counts over 7,000 residents within a 300 m walk.</p>
<p>Sixteen configurations from Qwen3, Gemma 3, Gemma 4 and Llama 3.x, from 1B to 14B parameters, are each run ten times per case: 1,440 responses and about 22,000 labelled claims.</p>
</div><div class="fusion-title title fusion-title-3 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Main findings</span></h2></div><div class="fusion-text fusion-text-4 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><ul>
<li>Family matters more than size. Out of 30, Gemma 4 scores 22.5 to 23 on trap resistance and Qwen3 10 to 20.5. Gemma 3 scores 1 to 4 and Llama 2 to 6.5. The 1.7B Qwen3 outscores the 12B Gemma 3.</li>
<li>Reading and defending are different skills. Several models report every figure correctly, then drop them the moment the user disagrees.</li>
<li>The more plausible the premise, the weaker the resistance: 65% in Chicago, 43% in Paris, 14% in Hanoi.</li>
<li>Thinking mode makes answers 2.6 to 4.6 times slower without a consistent gain in grounding.</li>
<li>Outcomes change between identical runs, so a single generation is one draw, not a verdict.</li>
</ul>
</div><iframe id="nscr-profiles"
  src="https://urbangeoanalytics.com/wp-content/uploads/2026/10/nscr-llm-model-profiles.html"
  title="Six-axis profile per model: grounding, hallucination, trap resistance, stability, concision and speed"
  style="width:100%;height:1040px;border:0;" loading="lazy"></iframe>
<script>
window.addEventListener('message', function (e) {
  if (e.data && e.data.nscrFigureHeight) {
    document.getElementById('nscr-profiles').style.height = (e.data.nscrFigureHeight + 10) + 'px';
  }
});
</script>
<p style="font-size:13px;text-align:center;">Six-axis profile per model. Use the buttons to switch between pooled and per-city profiles; hover a point for its value.</p><div class="fusion-title title fusion-title-4 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Why it matters</span></h2></div><div class="fusion-text fusion-text-5 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><div data-line="37" data-line-type="context" data-line-index="36">
<div data-line="3" data-line-type="context" data-line-index="2">Scoring answers only against the truth would rate many of these models as reliable. For decision-support uses, the practical lesson is that a larger model that yields to its user is a worse choice than a smaller one that holds to the data.</div>
<div data-line="4" data-line-type="context" data-line-index="3"></div>
<div data-line="5" data-line-type="context" data-line-index="4">This matters most in geography, where what a model believes about a place is often a reputation rather than a fact: the Marais as a quiet historic quarter, Hanoi&#8217;s Old Quarter as shops and tourists rather than residents. Planners, analysts and public agencies are starting to put local data in front of these models precisely to get past such assumptions. If the model drops that data as soon as a confident user repeats the stereotype, the retrieval step has achieved nothing.</div>
<div data-line="6" data-line-type="context" data-line-index="5"></div>
<div data-line="7" data-line-type="context" data-line-index="6">The paper also shows that this risk cannot be read off a model card. Parameter count does not predict it, a reasoning mode does not fix it, and one good answer does not rule it out. It has to be measured, on the data and the questions the model will actually face.</div>
</div>
</div><div class="fusion-title title fusion-title-5 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">References and Links</span></h2></div><div class="fusion-text fusion-text-6 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><ul>
<li>Paper (preprint): <a href="https://arxiv.org/abs/2609.39437">https://arxiv.org/abs/2609.39437</a></li>
<li>Code and data: <a href="https://github.com/perezjoan/NSCR-LLM">https://github.com/perezjoan/NSCR-LLM</a></li>
<li>Archived release v1.0.0: <a href="https://doi.org/10.5281/zenodo.23056312">https://doi.org/10.5281/zenodo.23056312</a></li>
</ul>
</div></div></div><div class="fusion-layout-column fusion_builder_column fusion-builder-column-1 awb-sticky awb-sticky-medium awb-sticky-large fusion_builder_column_1_4 1_4 fusion-flex-column" style="--awb-padding-top:20px;--awb-padding-right:20px;--awb-padding-bottom:20px;--awb-padding-left:20px;--awb-bg-size:cover;--awb-border-color:var(--awb-color6);--awb-border-style:solid;--awb-width-large:25%;--awb-margin-top-large:0px;--awb-spacing-right-large:7.68%;--awb-margin-bottom-large:20px;--awb-spacing-left-large:7.68%;--awb-width-medium:25%;--awb-order-medium:0;--awb-spacing-right-medium:7.68%;--awb-spacing-left-medium:7.68%;--awb-width-small:100%;--awb-order-small:0;--awb-spacing-right-small:1.92%;--awb-spacing-left-small:1.92%;--awb-sticky-offset:150px;" data-scroll-devices="small-visibility,medium-visibility,large-visibility"><div class="fusion-column-wrapper fusion-column-has-shadow fusion-flex-justify-content-flex-start fusion-content-layout-column"><div class="fusion-text fusion-text-7"><p><span style="color: #143c4e;"><strong>Table of contents</strong></span></p>
</div><div class="awb-toc-el awb-toc-el--1" data-awb-toc-id="1" data-awb-toc-options="{&quot;allowed_heading_tags&quot;:{&quot;h2&quot;:0},&quot;ignore_headings&quot;:&quot;&quot;,&quot;ignore_headings_words&quot;:&quot;&quot;,&quot;enable_cache&quot;:&quot;no&quot;,&quot;highlight_current_heading&quot;:&quot;yes&quot;,&quot;hide_hidden_titles&quot;:&quot;no&quot;,&quot;limit_container&quot;:&quot;page_content&quot;,&quot;select_custom_headings&quot;:&quot;.contenu H2, .contenu H3&quot;,&quot;icon&quot;:&quot;fa-flag fas&quot;,&quot;counter_type&quot;:&quot;none&quot;}" style="--awb-item-padding-right:5px;--awb-item-padding-left:5px;"><div class="awb-toc-el__content"></div></div><div class="fusion-separator fusion-full-width-sep" style="align-self: center;margin-left: auto;margin-right: auto;margin-top:20px;margin-bottom:20px;width:100%;"><div class="fusion-separator-border sep-single sep-solid" style="--awb-height:20px;--awb-amount:20px;--awb-sep-color:var(--awb-color6);border-color:var(--awb-color6);border-top-width:1px;"></div></div><div class="fusion-image-element " style="--awb-margin-top:25px;--awb-margin-bottom:25px;--awb-caption-title-font-family:var(--h2_typography-font-family);--awb-caption-title-font-weight:var(--h2_typography-font-weight);--awb-caption-title-font-style:var(--h2_typography-font-style);--awb-caption-title-size:var(--h2_typography-font-size);--awb-caption-title-transform:var(--h2_typography-text-transform);--awb-caption-title-line-height:var(--h2_typography-line-height);--awb-caption-title-letter-spacing:var(--h2_typography-letter-spacing);--awb-filter:saturate(100%);--awb-filter-transition:filter 0.3s ease;--awb-filter-hover:saturate(0%);"><span class=" fusion-imageframe imageframe-none imageframe-2 hover-type-zoomout"><img decoding="async" width="1536" height="1024" alt="Difficulty level 3: Advanced" src="https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3.png" class="img-responsive wp-image-1688" srcset="https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-200x133.png 200w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-400x267.png 400w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-600x400.png 600w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-800x533.png 800w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-1200x800.png 1200w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3.png 1536w" sizes="(max-width: 640px) 100vw, 400px" /></span></div></div></div></div></div>
<p>The post <a href="https://urbangeoanalytics.com/llm-faithfulness-benchmark-geographic-reasoning/">Do Open-Weight LLMs Reason From the Spatial Context They Are Given?</a> appeared first on <a href="https://urbangeoanalytics.com">Urban Geo Analytics</a>.</p>
]]></content:encoded>
					
		
		
			</item>
		<item>
		<title>UVLM v4.0.0 — Gemma 4, the Transformers 5 Migration, and Why This One Is a Major Version</title>
		<link>https://urbangeoanalytics.com/uvlm-4-0-0-gemma-4-transformers-5/</link>
					<comments>https://urbangeoanalytics.com/uvlm-4-0-0-gemma-4-transformers-5/#respond</comments>
		
		<dc:creator><![CDATA[Joan Perez]]></dc:creator>
		<pubDate>Mon, 17 Aug 2026 07:45:10 +0000</pubDate>
				<category><![CDATA[Advanced]]></category>
		<category><![CDATA[Package]]></category>
		<category><![CDATA[Python]]></category>
		<category><![CDATA[Vision Language Model]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[gemma]]></category>
		<category><![CDATA[Image Analysis]]></category>
		<category><![CDATA[Open Source]]></category>
		<category><![CDATA[UVLM]]></category>
		<guid isPermaLink="false">https://urbangeoanalytics.com/?p=3029</guid>

					<description><![CDATA[<p>Highlights  New model family: Gemma 4 (Google DeepMind, released April 2026) joins as the fifth family — E2B, E4B, and 12B Instruct, bringing the registry to 24 checkpoints Breaking change, done honestly: UVLM now requires Transformers ≥ 5.15; all four existing families were re-validated on GPU before release, and v3.2.0 remains installable  [...]</p>
<p>The post <a href="https://urbangeoanalytics.com/uvlm-4-0-0-gemma-4-transformers-5/">UVLM v4.0.0 — Gemma 4, the Transformers 5 Migration, and Why This One Is a Major Version</a> appeared first on <a href="https://urbangeoanalytics.com">Urban Geo Analytics</a>.</p>
]]></description>
										<content:encoded><![CDATA[<p><div class="fusion-fullwidth fullwidth-box fusion-builder-row-2 fusion-flex-container has-pattern-background has-mask-background nonhundred-percent-fullwidth non-hundred-percent-height-scrolling" style="--awb-border-radius-top-left:0px;--awb-border-radius-top-right:0px;--awb-border-radius-bottom-right:0px;--awb-border-radius-bottom-left:0px;--awb-flex-wrap:wrap;" id="contenu" ><div class="fusion-builder-row fusion-row fusion-flex-align-items-flex-start fusion-flex-content-wrap" style="max-width:1248px;margin-left: calc(-4% / 2 );margin-right: calc(-4% / 2 );"><div class="fusion-layout-column fusion_builder_column fusion-builder-column-2 fusion_builder_column_1_1 1_1 fusion-flex-column" style="--awb-bg-size:cover;--awb-width-large:100%;--awb-margin-top-large:0px;--awb-spacing-right-large:1.92%;--awb-margin-bottom-large:20px;--awb-spacing-left-large:1.92%;--awb-width-medium:100%;--awb-order-medium:0;--awb-spacing-right-medium:1.92%;--awb-spacing-left-medium:1.92%;--awb-width-small:100%;--awb-order-small:0;--awb-spacing-right-small:1.92%;--awb-spacing-left-small:1.92%;"><div class="fusion-column-wrapper fusion-column-has-shadow fusion-flex-justify-content-flex-start fusion-content-layout-column"><div class="fusion-image-element " style="--awb-caption-title-font-family:var(--h2_typography-font-family);--awb-caption-title-font-weight:var(--h2_typography-font-weight);--awb-caption-title-font-style:var(--h2_typography-font-style);--awb-caption-title-size:var(--h2_typography-font-size);--awb-caption-title-transform:var(--h2_typography-text-transform);--awb-caption-title-line-height:var(--h2_typography-line-height);--awb-caption-title-letter-spacing:var(--h2_typography-letter-spacing);"><span class=" fusion-imageframe imageframe-none imageframe-3 hover-type-none"><img decoding="async" width="1536" height="1024" alt="UVLM v4.0.0 — Gemma 4, the Transformers 5 Migration, and Why This One Is a Major Version" title="UVLM 4.0.0" src="https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0.png" class="img-responsive wp-image-3043" srcset="https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0-200x133.png 200w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0-400x267.png 400w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0-600x400.png 600w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0-800x533.png 800w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0-1200x800.png 1200w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/UVLM-4.0.0.png 1536w" sizes="(max-width: 640px) 100vw, 1200px" /></span></div></div></div><div class="fusion-layout-column fusion_builder_column fusion-builder-column-3 fusion_builder_column_3_4 3_4 fusion-flex-column" style="--awb-bg-size:cover;--awb-width-large:75%;--awb-margin-top-large:0px;--awb-spacing-right-large:2.56%;--awb-margin-bottom-large:20px;--awb-spacing-left-large:2.56%;--awb-width-medium:75%;--awb-order-medium:0;--awb-spacing-right-medium:2.56%;--awb-spacing-left-medium:2.56%;--awb-width-small:100%;--awb-order-small:0;--awb-spacing-right-small:1.92%;--awb-spacing-left-small:1.92%;" id="contenu" data-scroll-devices="small-visibility,medium-visibility,large-visibility"><div class="fusion-column-wrapper fusion-column-has-shadow fusion-flex-justify-content-flex-start fusion-content-layout-column"><div class="fusion-text fusion-text-8 fusion-text-no-margin" style="--awb-margin-bottom:-10px;"><h5><strong>Highlights</strong></h5>
</div><div class="fusion-text fusion-text-9" style="--awb-margin-top:-20px;"><ul>
<li><strong>New model family:</strong> Gemma 4 (Google DeepMind, released April 2026) joins as the fifth family — E2B, E4B, and 12B Instruct, bringing the registry to 24 checkpoints</li>
<li><strong>Breaking change, done honestly:</strong> UVLM now requires Transformers ≥ 5.15; all four existing families were re-validated on GPU before release, and <a href="https://urbangeoanalytics.com/uvlm-3-2-0-internvl-backend/">v3.2.0</a> remains installable for Transformers 4.x environments</li>
<li><strong>Measured, not guessed:</strong> Gemma 4&#8217;s &#8220;effective parameters&#8221; hide ~10–16 GB raw checkpoints — this release documents exactly what runs on an 8 GB GPU, and how</li>
</ul>
</div><div class="fusion-title title fusion-title-6 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Why a major version?</span></h2></div><div class="fusion-text fusion-text-10 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p dir="ltr"><a href="https://github.com/perezjoan/UVLM">UVLM</a> has followed one rule since the package release: a new model family is a minor version, because it breaks nothing. v4.0.0 breaks that streak for a reason we could not code around: <strong>Gemma 4 does not exist in any Transformers 4.x release.</strong> We verified this empirically — 4.57.6 is the final version of the 4.x line, and it does not register the <code class="bg-text-200/5 border border-0.5 border-border-300 text-danger-000 rounded-&#091;0.4rem&#093; px-1 py-px text-&#091;0.9rem&#093;">gemma4</code> architecture; support begins in the 5.x line. Adopting the family therefore means lifting the <code class="bg-text-200/5 border border-0.5 border-border-300 text-danger-000 rounded-&#091;0.4rem&#093; px-1 py-px text-&#091;0.9rem&#093;">transformers &lt; 5.0.0</code> cap that UVLM has carried since v3.0.1.</p>
<p dir="ltr">That cap was not decoration. It existed because early Transformers 5.x releases crashed Qwen2.5-VL at load time with a weight-conversion error. So before this release, all four existing families — LLaVA-NeXT, Qwen2.5-VL, Qwen3-VL, InternVL3.5 — were re-validated on GPU under Transformers 5.15, in 4-bit, on real inference tasks. The historical Qwen2.5-VL crash is <strong>confirmed fixed</strong>: the 7B model loads and answers correctly at full speed. That validation is what makes this a release rather than a gamble.</p>
<p dir="ltr">If your environment must stay on Transformers 4.x, nothing is taken from you: <code class="bg-text-200/5 border border-0.5 border-border-300 text-danger-000 rounded-&#091;0.4rem&#093; px-1 py-px text-&#091;0.9rem&#093;">pip install git+https://github.com/perezjoan/UVLM.git@v3.2.0</code> pins the last 4.x-compatible release, permanently.</p>
</div><div class="fusion-title title fusion-title-7 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">What is Gemma 4?</span></h2></div><div class="fusion-text fusion-text-11 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>Gemma 4 is Google DeepMind&#8217;s latest open multimodal generation, released in April 2026 under Apache 2.0. UVLM v4.0.0 integrates three Instruct checkpoints:</p>
</div>
<div class="table-1">
<table width="100%">
<thead>
<tr>
<th align="left">Model</th>
<th align="left"> Parameters</th>
<th align="left"> Raw checkpoint</th>
<th align="left"> Runs on</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Gemma 4 E2B Instruct</td>
<td align="left"> ~2B effective</td>
<td align="left"> ~10 GB</td>
<td align="left"> 8 GB GPU in FP16 with CPU offload</td>
</tr>
<tr>
<td align="left">Gemma 4 E4B Instruct</td>
<td align="left"> ~4B effective</td>
<td align="left"> ~16 GB</td>
<td align="left">Larger-VRAM environments (Colab A100/L4)</td>
</tr>
<tr>
<td align="left">Gemma 4 12B Instruct</td>
<td align="left"> 12B</td>
<td align="left"> ~24 GB</td>
<td align="left">Larger-VRAM environments (Colab A100/L4)</td>
</tr>
</tbody>
</table>
</div>
<div class="fusion-text fusion-text-12 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p dir="ltr">Notice the third column, because it is this release&#8217;s most useful finding. E2B and E4B are <strong>&#8220;effective&#8221;-parameter models</strong>: Per-Layer Embeddings give them the <em>compute</em> profile of a 2B/4B model, but the embedding tables push the <em>raw</em> checkpoint far beyond what the name suggests. A &#8220;2B&#8221; model that downloads 10 GB of weights behaves very differently from Qwen3-VL 2B&#8217;s genuinely small footprint — and honest benchmarking infrastructure should say so, with numbers.</p>
<p dir="ltr">Technically, Gemma 4 shares the tokenizing-chat-template pipeline introduced with InternVL3.5, with one addition: Gemma 4 models other than E2B/E4B wrap their output in thought-channel tags even when thinking is disabled, and the backend strips them automatically. As always, the family appeared in the notebook selector with zero interface changes.</p>
</div><div class="fusion-title title fusion-title-8 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">What actually runs on a laptop GPU</span></h2></div><div class="fusion-text fusion-text-13 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>We validated Gemma 4 on an 8 GB RTX 5060, and the result inverts the usual intuition: <strong>FP16 is the low-memory mode.</strong> In FP16, the compute-heavy layers stay on the GPU while the PLE embedding tables — lookup-only structures designed to live off-accelerator — offload to system RAM in half precision. Measured: about 17 seconds per image. Slow, but fully functional.</p>
</div><div class="fusion-image-element " style="--awb-caption-title-font-family:var(--h2_typography-font-family);--awb-caption-title-font-weight:var(--h2_typography-font-weight);--awb-caption-title-font-style:var(--h2_typography-font-style);--awb-caption-title-size:var(--h2_typography-font-size);--awb-caption-title-transform:var(--h2_typography-text-transform);--awb-caption-title-line-height:var(--h2_typography-line-height);--awb-caption-title-letter-spacing:var(--h2_typography-letter-spacing);"><span class=" fusion-imageframe imageframe-none imageframe-4 hover-type-none"><img decoding="async" width="1215" height="641" alt="Gemma 4 joins UVLM as the fifth vision-language model family" title="gemma illustration" src="https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration.png" class="img-responsive wp-image-3039" srcset="https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration-200x106.png 200w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration-400x211.png 400w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration-600x317.png 600w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration-800x422.png 800w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration-1200x633.png 1200w, https://urbangeoanalytics.com/wp-content/uploads/2026/08/gemma-illustration.png 1215w" sizes="(max-width: 640px) 100vw, 1200px" /></span></div><div class="fusion-text fusion-text-14 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>4-bit quantization — normally the memory-saver — fails here, for a subtle reason: offloaded modules are kept in FP32, roughly doubling the RAM requirement, and any further spill to disk is unsupported by bitsandbytes. Rather than leave users with a 200-line traceback, v4.0.0 detects this case and raises a two-sentence error recommending FP16. To support all of this, the loader gained general CPU-offload capability for oversized checkpoints — a change that only <em>permits</em> offload: models that fit entirely on the GPU are placed exactly as before.</p>
</div><div class="fusion-title title fusion-title-9 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Also in this release</span></h2></div><div class="fusion-text fusion-text-15 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p>A failed model load in the notebooks now invalidates the previously loaded model, so a batch run after a failed load errors out loudly instead of silently benchmarking the wrong checkpoint — a trap we fell into ourselves during validation, and one that the per-model output filenames from v3.2.0 caught. InternVL3.5 users on Transformers 5 will see a harmless &#8220;tied weights&#8221; warning caused by an upstream config inconsistency; Transformers resolves it correctly.</p>
</div><div class="fusion-title title fusion-title-10 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Getting started</span></h2></div><div class="fusion-text fusion-text-16" style="--awb-margin-top:25px;"><pre class="EnlighterJSRAW" data-enlighter-language="bash" data-enlighter-theme="dracula" data-enlighter-group="Bash" data-enlighter-title="Bash">pip install --upgrade --force-reinstall git+https://github.com/perezjoan/UVLM.git</pre>
</div><div class="fusion-text fusion-text-17 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-bottom:25px;"><p>Note: no <code class="bg-text-200/5 border border-0.5 border-border-300 text-danger-000 rounded-&#091;0.4rem&#093; px-1 py-px text-&#091;0.9rem&#093;">--no-deps</code> this time — the whole point of the upgrade is that pip pulls Transformers 5.15 for you. Colab users get v4.0.0 automatically on their next session, since the notebook always installs the latest version; the re-validation above is what makes that automatic jump safe. The three-block workflow, consensus validation, chain-of-thought mode, and truncation detection all work with Gemma 4 out of the box.</p>
</div><div class="fusion-title title fusion-title-11 fusion-sep-none fusion-title-text fusion-title-size-two" style="--awb-margin-top:25px;--awb-margin-bottom:0px;--awb-font-size:35px;"><h2 class="fusion-title-heading title-heading-left fusion-responsive-typography-calculated" style="margin:0;font-size:1em;--fontSize:35;line-height:var(--awb-typography1-line-height);"><span style="font-weight: 400;">Where UVLM stands</span></h2></div><div class="fusion-text fusion-text-18 fusion-text-no-margin" style="--awb-content-alignment:justify;--awb-margin-top:25px;--awb-margin-bottom:25px;"><p dir="ltr">Five months ago, UVLM was a two-family package. It now supports <strong>five families and 24 checkpoints from 1B to 110B parameters</strong> — LLaVA-NeXT, Qwen2.5-VL, Qwen3-VL, InternVL3.5, Gemma 4 — behind one interface, one prompt format, one evaluation protocol. Three families were added in three releases without a single notebook edit, each validated on hardware before shipping. That is the registry we will be benchmarking against in upcoming applied work — more on that soon.</p>
<p dir="ltr">Full change log in <a href="https://github.com/perezjoan/UVLM/blob/main/VERSIONS.txt">VERSIONS.txt</a> · Source and releases on GitHub · If you use UVLM in research, please cite our software paper.</p>
</div></div></div><div class="fusion-layout-column fusion_builder_column fusion-builder-column-4 awb-sticky awb-sticky-medium awb-sticky-large fusion_builder_column_1_4 1_4 fusion-flex-column" style="--awb-padding-top:20px;--awb-padding-right:20px;--awb-padding-bottom:20px;--awb-padding-left:20px;--awb-bg-size:cover;--awb-border-color:var(--awb-color6);--awb-border-style:solid;--awb-width-large:25%;--awb-margin-top-large:0px;--awb-spacing-right-large:7.68%;--awb-margin-bottom-large:20px;--awb-spacing-left-large:7.68%;--awb-width-medium:25%;--awb-order-medium:0;--awb-spacing-right-medium:7.68%;--awb-spacing-left-medium:7.68%;--awb-width-small:100%;--awb-order-small:0;--awb-spacing-right-small:1.92%;--awb-spacing-left-small:1.92%;--awb-sticky-offset:150px;" data-scroll-devices="small-visibility,medium-visibility,large-visibility"><div class="fusion-column-wrapper fusion-column-has-shadow fusion-flex-justify-content-flex-start fusion-content-layout-column"><div class="fusion-text fusion-text-19"><p><span style="color: #143c4e;"><strong>Table of contents</strong></span></p>
</div><div class="awb-toc-el awb-toc-el--2" data-awb-toc-id="2" data-awb-toc-options="{&quot;allowed_heading_tags&quot;:{&quot;h2&quot;:0},&quot;ignore_headings&quot;:&quot;&quot;,&quot;ignore_headings_words&quot;:&quot;&quot;,&quot;enable_cache&quot;:&quot;no&quot;,&quot;highlight_current_heading&quot;:&quot;yes&quot;,&quot;hide_hidden_titles&quot;:&quot;no&quot;,&quot;limit_container&quot;:&quot;page_content&quot;,&quot;select_custom_headings&quot;:&quot;.contenu H2, .contenu H3&quot;,&quot;icon&quot;:&quot;fa-flag fas&quot;,&quot;counter_type&quot;:&quot;none&quot;}" style="--awb-item-padding-right:5px;--awb-item-padding-left:5px;"><div class="awb-toc-el__content"></div></div><div class="fusion-separator fusion-full-width-sep" style="align-self: center;margin-left: auto;margin-right: auto;margin-top:20px;margin-bottom:20px;width:100%;"><div class="fusion-separator-border sep-single sep-solid" style="--awb-height:20px;--awb-amount:20px;--awb-sep-color:var(--awb-color6);border-color:var(--awb-color6);border-top-width:1px;"></div></div><div class="fusion-image-element " style="--awb-margin-top:25px;--awb-margin-bottom:25px;--awb-caption-title-font-family:var(--h2_typography-font-family);--awb-caption-title-font-weight:var(--h2_typography-font-weight);--awb-caption-title-font-style:var(--h2_typography-font-style);--awb-caption-title-size:var(--h2_typography-font-size);--awb-caption-title-transform:var(--h2_typography-text-transform);--awb-caption-title-line-height:var(--h2_typography-line-height);--awb-caption-title-letter-spacing:var(--h2_typography-letter-spacing);--awb-filter:saturate(100%);--awb-filter-transition:filter 0.3s ease;--awb-filter-hover:saturate(0%);"><span class=" fusion-imageframe imageframe-none imageframe-5 hover-type-zoomout"><img decoding="async" width="1536" height="1024" alt="Difficulty level 3: Advanced" src="https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3.png" class="img-responsive wp-image-1688" srcset="https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-200x133.png 200w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-400x267.png 400w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-600x400.png 600w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-800x533.png 800w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3-1200x800.png 1200w, https://urbangeoanalytics.com/wp-content/uploads/2025/11/blog-lvl3.png 1536w" sizes="(max-width: 640px) 100vw, 400px" /></span></div></div></div></div></div><div class="fusion-fullwidth fullwidth-box fusion-builder-row-3 fusion-flex-container has-pattern-background has-mask-background nonhundred-percent-fullwidth non-hundred-percent-height-scrolling" style="--awb-border-radius-top-left:0px;--awb-border-radius-top-right:0px;--awb-border-radius-bottom-right:0px;--awb-border-radius-bottom-left:0px;--awb-flex-wrap:wrap;" ><div class="fusion-builder-row fusion-row fusion-flex-align-items-flex-start fusion-flex-content-wrap" style="max-width:1248px;margin-left: calc(-4% / 2 );margin-right: calc(-4% / 2 );"></div></div></p>
<p>The post <a href="https://urbangeoanalytics.com/uvlm-4-0-0-gemma-4-transformers-5/">UVLM v4.0.0 — Gemma 4, the Transformers 5 Migration, and Why This One Is a Major Version</a> appeared first on <a href="https://urbangeoanalytics.com">Urban Geo Analytics</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://urbangeoanalytics.com/uvlm-4-0-0-gemma-4-transformers-5/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
			</item>
	</channel>
</rss>
