Which Ollama Cloud Model is Best? Duck Driving a Jeep SVG Comparison (13 Models)

If you are wondering which Ollama cloud model is best for SVG generation, this post is for you. We sent the exact same prompt to 13 different state-of-the-art LLMs and let them draw a duck driving a jeep. The results are wildly different, and they reveal which models are actually good at structured drawing tasks.

This is the third in our SVG benchmark series. The prompt was: Make an svg image about a duck driving a jeep. You can compare with our previous benchmarks: duck on a bicycle and duck jumping from a plane with a parachute.

The goal of this post is not to declare a winner – it is to give you the data so you can pick the best model for your own use case. Different models shine at different things: raw detail, clean structure, fast generation, or visual fidelity. We show you the SVG, the stats, and a short analysis for each. You decide.

How to Choose the Best Ollama Model for SVG

Before looking at the results, here are the criteria you should use to judge each model:

• Visual fidelity: Does the output actually look like a duck driving a jeep? Are the wheels round, the beak duck-like, and the proportions correct?
• Detail level: How much scene detail (windows, doors, terrain, sky) does the model add? More detail is not always better, but it shows effort.
• SVG code quality: Does it use <defs>, <use>, gradients, and filters, or just raw shapes? Cleaner code means easier to edit and reuse.
• Output size: Very small SVGs (under 2,000 chars) are usually too simple. Very large SVGs (over 15,000 chars) are rich but may be slow to render.
• Generation speed: If you are building an app that streams SVGs to users, faster matters. We report generation time for each model.
• Concept adherence: Did the model actually draw a jeep (with wheels, doors, windshield) and a duck (with a beak, body, eyes)? Some models lose the plot.

How It Works

The script discovers all cloud-hosted models via the Ollama API (/api/tags), pulls each model, then sends the identical prompt through the OpenAI-compatible endpoint (http://localhost:11434/v1/chat/completions). Each model’s response is parsed for an <svg>...</svg> block, and the extracted SVG is saved for rendering with zero post-processing.

Cloud models are identified by the remote_host field in the API response – these models are hosted on Ollama Cloud rather than running locally. This means even very large models (671B parameters) can be queried instantly without local GPU resources.

Summary Table: Compare All 13 Models at a Glance

Use this table to quickly compare models on the metrics that matter. The verdict column is a one-line summary to help you shortlist – but read the per-model sections below for the full picture before you decide.

13 out of 17 active models produced a valid SVG. The 4 retired models returned HTTP 410 Gone (removed from Ollama Cloud on 2026-07-15).

Quick Recommendation by Use Case

If you just want a shortcut, here is which model to pick based on what you care about:

• You want the most detailed, visually rich SVG: pick glm-5.1:cloud or nemotron-3-ultra:cloud
• You want the fastest response: pick gemma4:31b-cloud or gpt-oss:120b-cloud (under 12 seconds)
• You want the cleanest, most reusable SVG code: pick deepseek-v4-pro:cloud or kimi-k2.6:cloud
• You want a small, efficient SVG for web embedding: pick gemma4:cloud or nemotron-3-super:cloud
• You want a balance of detail and speed: pick deepseek-v4-flash:cloud or glm-5.2:cloud
• You want the newest model families: pick gpt-oss:120b-cloud (OpenAI open weights), minimax-m3:cloud, or gemma4:cloud

Now read on for the full per-model breakdown and judge for yourself.

1. deepseek-v4-flash_cloud

SVG size: 8244 characters
Complexity: Medium
Shape elements: 83
Distinct colors: 32
Raw response: 8671 characters

Analysis

This SVG contains approximately 83 shape elements and uses 32 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a moderately detailed SVG (5,000-10,000 characters), balancing visual richness with code clarity. The model included enough detail to make the scene recognizable while keeping the markup manageable. Good for general use.

2. deepseek-v4-pro_cloud

SVG size: 6639 characters
Complexity: Medium
Shape elements: 62
Distinct colors: 21
Raw response: 7062 characters

Analysis

This SVG contains approximately 62 shape elements and uses 21 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a moderately detailed SVG (5,000-10,000 characters), balancing visual richness with code clarity. The model included enough detail to make the scene recognizable while keeping the markup manageable. Good for general use.

3. gemma4_31b-cloud

SVG size: 1656 characters
Complexity: Low
Shape elements: 19
Distinct colors: 13
Raw response: 2326 characters

Analysis

This SVG contains approximately 19 shape elements and uses 13 distinct colors.

Notable SVG techniques used:

• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

4. gemma4_cloud

SVG size: 1694 characters
Complexity: Low
Shape elements: 16
Distinct colors: 13
Raw response: 2644 characters

Analysis

This SVG contains approximately 16 shape elements and uses 13 distinct colors.

Notable SVG techniques used:

• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

5. glm-5.1_cloud

SVG size: 17880 characters
Complexity: Very high
Shape elements: 160
Distinct colors: 50
Raw response: 21943 characters

Analysis

This SVG contains approximately 160 shape elements and uses 50 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Adds <text> labels, showing the model tried to annotate the scene
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

With over 15,000 characters of SVG markup, this is one of the most detailed outputs in the comparison. The model invested significant effort in adding fine details like the jeep body panels, terrain, sky, and the duck’s accessories. Best for users who want maximum visual richness.

6. glm-5.2_cloud

SVG size: 6971 characters
Complexity: Medium
Shape elements: 66
Distinct colors: 26
Raw response: 7034 characters

Analysis

This SVG contains approximately 66 shape elements and uses 26 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a moderately detailed SVG (5,000-10,000 characters), balancing visual richness with code clarity. The model included enough detail to make the scene recognizable while keeping the markup manageable. Good for general use.

7. gpt-oss_120b-cloud

SVG size: 2147 characters
Complexity: Low
Shape elements: 18
Distinct colors: 10
Raw response: 2917 characters

Analysis

This SVG contains approximately 18 shape elements and uses 10 distinct colors.

Notable SVG techniques used:

• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

8. kimi-k2.6_cloud

SVG size: 9941 characters
Complexity: Medium
Shape elements: 51
Distinct colors: 32
Raw response: 11079 characters

Analysis

This SVG contains approximately 51 shape elements and uses 32 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Applies SVG filters (e.g., drop shadows, blurs) for visual effects
• Uses transforms (translate, rotate, scale) for precise positioning
• Includes animation elements – the model attempted motion effects
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a moderately detailed SVG (5,000-10,000 characters), balancing visual richness with code clarity. The model included enough detail to make the scene recognizable while keeping the markup manageable. Good for general use.

9. minimax-m2.7_cloud

SVG size: 2697 characters
Complexity: Low
Shape elements: 26
Distinct colors: 15
Raw response: 2904 characters

Analysis

This SVG contains approximately 26 shape elements and uses 15 distinct colors.

Notable SVG techniques used:

• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

10. minimax-m3_cloud

SVG size: 6464 characters
Complexity: Medium
Shape elements: 74
Distinct colors: 26
Raw response: 7444 characters

Analysis

This SVG contains approximately 74 shape elements and uses 26 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a moderately detailed SVG (5,000-10,000 characters), balancing visual richness with code clarity. The model included enough detail to make the scene recognizable while keeping the markup manageable. Good for general use.

11. nemotron-3-super_cloud

SVG size: 1782 characters
Complexity: Low
Shape elements: 15
Distinct colors: 10
Raw response: 2447 characters

Analysis

This SVG contains approximately 15 shape elements and uses 10 distinct colors.

Notable SVG techniques used:

• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

12. nemotron-3-ultra_cloud

SVG size: 16049 characters
Complexity: Very high
Shape elements: 92
Distinct colors: 40
Raw response: 17462 characters

Analysis

This SVG contains approximately 92 shape elements and uses 40 distinct colors.

Notable SVG techniques used:

• Uses <defs> blocks for reusable components (a sign of well-structured, editable SVG)
• Includes gradient fills for richer visual depth and shading
• Applies SVG filters (e.g., drop shadows, blurs) for visual effects
• Uses transforms (translate, rotate, scale) for precise positioning
• Includes animation elements – the model attempted motion effects
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

With over 15,000 characters of SVG markup, this is one of the most detailed outputs in the comparison. The model invested significant effort in adding fine details like the jeep body panels, terrain, sky, and the duck’s accessories. Best for users who want maximum visual richness.

13. qwen3.5_397b-cloud

SVG size: 3081 characters
Complexity: Low
Shape elements: 29
Distinct colors: 16
Raw response: 3142 characters

Analysis

This SVG contains approximately 29 shape elements and uses 16 distinct colors.

Notable SVG techniques used:

• Uses transforms (translate, rotate, scale) for precise positioning
• Contains XML comments, showing the model’s internal documentation of its work

Concept adherence:

• Duck is present (the model understood the main subject)
• Jeep / vehicle elements are present (the model understood the second subject)

This is a compact SVG (under 5,000 characters) that uses minimal markup to convey the scene. The model prioritized efficiency and simplicity over elaborate detail. Best for embedding in web pages where small file size matters.

Retired Models

The following models were listed in the Ollama registry but returned HTTP 410 Gone when queried. They were retired from Ollama Cloud and are no longer available for inference:

• bjoernb/claude-opus-4-5:latest – glm-4.7 was retired at 2026-07-15
• deepseek-v3.1:671b-cloud – deepseek-v3.1:671b was retired at 2026-07-15
• glm-5:cloud – glm-5 was retired at 2026-07-15
• qwen3-vl:235b-cloud – qwen3-vl:235b was retired at 2026-06-16

How to Pick the Best Model: A Decision Guide

Now that you have seen all 13 SVGs, here is a structured way to decide which Ollama cloud model is best for your SVG generation needs:

Step 1: Define your priority

• Visual quality first: Scroll back through the SVGs above and pick the one that looks best to your eye. There is no substitute for visual judgment. The shape and color counts are useful, but a model with 60 shapes can look worse than one with 20 if the composition is off.
• Code quality first: Open the raw SVG source for each model (use the disclosure toggles) and look for <defs>, <use>, gradients, and clean indentation. Models that produce structured code are easier to edit and reuse programmatically.
• Speed first: If you are building a real-time app, prioritize the models that responded in under 15 seconds (gemma4:31b-cloud, gemma4:cloud, gpt-oss:120b-cloud, nemotron-3-super:cloud).
• File size first: For web embedding, smaller is better. Look at the SVG size column in the summary table.

Step 2: Cross-check across prompts

A model that does well on one prompt might fail on another. Check our other benchmarks:

• Duck on a bicycle – simpler scene
• Duck with a parachute – dynamic action scene
• This post (duck in a jeep) – vehicle with multiple parts

A model that consistently produces good results across all three prompts is a safer pick than one that only shines occasionally.

Step 3: Test with your own prompt

Every model has strengths and weaknesses. The only way to know for sure which model is best for your specific use case is to test it with your own prompt. The Ollama Cloud API is OpenAI-compatible, so you can use any standard client:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
resp = client.chat.completions.create(
    model="deepseek-v4-pro:cloud",  # change this to test different models
    messages=[{"role": "user", "content": "Make an svg image of <your prompt>"}],
)
print(resp.choices[0].message.content)

New Models Added in This Update

This update added 4 new cloud models that were not in the original comparison:

• deepseek-v4-flash:cloud – a faster variant of deepseek-v4-pro, balancing speed and detail
• gemma4:cloud – the standard Gemma 4 variant (alongside the 31B version)
• gpt-oss:120b-cloud – OpenAI’s open-weight 120B model, available on Ollama Cloud
• minimax-m3:cloud – the latest MiniMax M3 model (alongside M2.7)

These new models let you compare within the same family (e.g. deepseek-v4-pro vs deepseek-v4-flash, minimax-m2.7 vs minimax-m3, gemma4 vs gemma4:31b) to see how different sizes and variants perform on the same prompt.

Conclusion: You Decide the Winner

This comparison shows that 13 out of 17 active Ollama cloud models can generate valid SVG artwork from a natural language prompt involving a vehicle (jeep) with multiple parts. The results vary dramatically in complexity, style, and technique – and there is no single “best” model.

Our takeaways after running three SVG benchmarks (bicycle, parachute, jeep):

• deepseek-v4-pro:cloud consistently produces well-structured, technically advanced SVGs with <defs>, <use>, and transforms. A strong default choice for code quality.
• glm-5.1:cloud and nemotron-3-ultra:cloud consistently produce the longest, most detailed SVGs. Best when you want maximum visual richness.
• gemma4:31b-cloud and gpt-oss:120b-cloud are consistently among the fastest (under 12 seconds) and produce simple, compact SVGs. Best for speed-sensitive applications.
• kimi-k2.6:cloud often uses animations, which is unique among the models – but the animations sometimes make content invisible unless you fix them (as we did in the parachute post).
• glm-5.2:cloud and deepseek-v4-flash:cloud offer a reliable balance of detail, speed, and code quality.

But the real verdict is yours. Scroll back through the SVGs, compare them visually, check the raw code, and pick the model that best fits your needs. Every model in this comparison is available right now on Ollama Cloud – so you can reproduce these results in minutes.

Links

• Previous: Duck Driving a Bicycle Comparison
• Previous: Duck Jumping From a Plane Comparison
• Ollama Official Website
• Ollama Cloud Documentation
• SVG Specification (MDN)
• OpenAI API Reference (used by Ollama)

Enjoyed this post? Never miss out on future posts by following us