First, the distilled models from DeepSeek aren’t actually DeepSeek, and second, the performance of the few models that can be deployed on consumer-grade graphics cards is indeed quite mediocre. In comparison, it’s not as good as using the API.<\/p>\n
So, entering today’s topic, let’s see how Qwen3’s small models actually perform.<\/p>\n
First, a disclaimer: My testing scenarios this time are quite subjective and not a serious review. The relevant scores are given by AI (Doubao-1.5-thinking-pro) for entertainment.<\/p>\n
In terms of scenarios, I selected three based on my own usage. I think a significant proportion of ordinary office workers have scenarios similar to mine:<\/p>\n
- \n
- Review and revision of purposeful copywriting.<\/li>\n<\/ol>\n
The test questions used 10 problematic positive review copies from Taobao.
\n2. Summarization and understanding of content.<\/p>\nThe test questions were 10 articles randomly selected from the internet.
\n3. Logical and computational abilities under basic Q&A.<\/p>\nThe test questions were 10 randomly selected questions (with answers) from the ‘Weak Intelligence Bar’ training set.<\/p>\n
A total of 10 large models participated in the test:<\/p>\n
- \n
- 6 were deployed locally via Ollama, namely:<\/li>\n<\/ul>\n
qwen3:8b, qwen3:14b, qwen3:32b, qwen3:30b-a3b, deepseek-r1:8b, deepseek-r1:32b
\n– 4 were accessed via online APIs, namely:<\/p>\nQwQ-32B\u3001GLM-4-Flash\u3001DeepSeek-V3\u3001Grok-3<\/p>\n
The judge is Doubao-1.5-thinking-pro:<\/p>\n
- \n
- For the first two scenarios, it will adopt a God’s-eye view, reviewing all contestants’ answers before scoring<\/li>\n
- For the Weak Intelligence Bar test, it will score based on the standard answers<\/li>\n<\/ul>\n
For the testing environment, I initially planned to use CherryStudio, but later found that models would copy each other’s homework when answering the same question. So, I ended up using a multi-dimensional table.<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/001-ada482f09ea3.png\")
<\/p>\n![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/002-1d2e574c627a.png\")
<\/p>\nThis also made it convenient to present the complete test questions and contestants’ answers at the end.<\/p>\n
Below are the official test results.<\/p>\n
Each scenario has 10 questions; the average score is the sum of scores divided by 10.<\/p>\n
Test 1: Copywriting Analysis and Revision:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/003-261f1db45ea9.png\")
<\/p>\nScores:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/004-0342e6db50bc.png\")
<\/p>\nQwen3 indeed performed impressively, with the 32B model taking the top spot in the first test.<\/p>\n
The two previously distilled DeepSeek models\u2014as I said, I wasn’t lying\u2014ranked relatively low.<\/p>\n
The freely callable GLM-4-Flash unfortunately ranked last. But to be fair to it, it’s hard to find an API that’s as free, unlimited, and high-concurrency as this one.<\/p>\n
Test 2: Article Summarization and Reflection:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/005-ea11c5b0546c.png\")
<\/p>\nScores:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/006-9deca7a7f80e.png\")
<\/p>\nQWQ-32B took the lead.<\/p>\n
DeepSeek-V3 secured second place.<\/p>\n
The previous first-place Qwen3:32B came in third this time.<\/p>\n
The distilled 32B from DeepSeek (actually distilled from Qwen) rose to fourth, while the 8B still lagged behind.<\/p>\n
GLM-4-Flash continued to trail along.<\/p>\n
Test 3: Weak Intelligence Bar Answer Exam:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/007-ffc5c0bdc2ab.png\")
<\/p>\nScores:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/008-88cc05662f20.png\")
<\/p>\nDeepSeek-V3 remained strong (this time using the 0324 version).<\/p>\n
Qwen3:14B made a bold move to secure second place, with 32B in third, also an excellent performer.<\/p>\n
The distilled 8B from DeepSeek unfortunately ranked last, with 32B third from the bottom.<\/p>\n
GLM-4-Flash made a comeback on the Weak Intelligence Bar, saving some face.<\/p>\n
Final Test Results:<\/p>\n
![\"\"](\"https:\/\/blog.liu-qi.cn\/wp-content\/uploads\/2025\/05\/009-edae7fdd777c.png\")
<\/p>\nDeepSeek-V3 took first place.<\/p>\n
Qwen3:32B came in second.<\/p>\n
QWQ-32B surprisingly secured third place.<\/p>\n
Qwen3:30B-A3B, which I had high hopes for, unexpectedly performed poorly in these tests.<\/p>\n
Based on these results, for friends with 90-series or higher GPUs, I recommend choosing Qwen3:32B for local deployment. The slightly inferior 14B is also a good choice, and the 8B will outperform the previous distilled DeepSeek versions.<\/p>\n
You can view the complete test questions and results at the link:<\/p>\n
https:\/\/ilovezhiwai.feishu.cn\/wiki\/HThDwnX0FiyTIakyDe9c2z99nef?from=from_copylink<\/p>\n
Let me reiterate: This test is neither scientific nor rigorous, and is purely for entertainment. Please do not use the conclusions from this test in serious contexts.<\/p>\n
Wishing everyone a happy May Day holiday!<\/p>\n","protected":false},"excerpt":{"rendered":"
A practical test evaluates Qwen3 small models for local deployment across three real-world scenarios, comparing 10 models with AI scoring to see if they’re viable for ordinary users.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[24],"tags":[21,20,10],"class_list":["post-141","post","type-post","status-publish","format-standard","hentry","category-articles","tag-ollama","tag-qwen","tag-10"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/141","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=141"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/141\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=141"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=141"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=141"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- 6 were deployed locally via Ollama, namely:<\/li>\n<\/ul>\n