Qwen vs Gemma vs DeepSeek: How the Top Open LLMs Compare in 2026

Qwen is the most popular open LLM family in 2026 by Hugging Face downloads. It held 44.2% of open LLM downloads in the week of Nov. 9, compared with 9.3% for Gemma and 4.3% for DeepSeek, according to Azumo’s Valkyrie research. Download share measures adoption rather than model quality, so technical teams should test candidates on their own workloads and data.

Written by:
November 20, 2026
Qwen vs Gemma vs DeepSeek

Qwen is the most downloaded open LLM (large language model) family on Hugging Face, the public hub for open AI models. In the week of Nov. 9, 2026, Qwen held 44.2% of open LLM downloads, ahead of Gemma at 9.3% and DeepSeek at 4.3%, according to our live LLM download trends.

This comparison comes from our own research. We have recorded Hugging Face's top 5,000 models by 30-day downloads every day since Feb. 12, 2026. In the weekly share, copies and fine-tunes that name their base model count toward that base's family. Downloads show adoption, not quality, so test any shortlist on your own data.

Key Statistics

Our daily tracking produced these headline figures for the three families. "Mid-March" means the week of March 15 to 21, 2026.

  1. In the week of Nov. 9, 2026, Qwen held 44.2% of open LLM downloads, Gemma 9.3% and DeepSeek 4.3%.
  2. The three families together held 57.8% of open LLM downloads in the week of Nov. 9, 2026, down from 71.1% in the week of May 11, 2026.
  3. In the 30 days to Nov. 12, 2026, Qwen placed 16 models in the top 25 open LLMs with 105.4M downloads combined. Gemma placed 3 with 16.5M, and DeepSeek placed 1 with 8.1M.
  4. Qwen's 44.2% in the week of Nov. 9, 2026 was its lowest weekly share since our weekly share series began in mid-March.
  5. Gemma's weekly share rose from 3.6% in the week of March 16, 2026 to a 15.6% peak in the week of July 13, 2026.
  6. DeepSeek-V3.2, at 685B parameters, was the largest model in the top 25 open LLMs by 30-day downloads as of Nov. 12, 2026.
  7. Apache-2.0 accounted for 62.6% and MIT for 18.9% of the most downloaded open LLMs on Nov. 12, 2026.
  8. Credited to the lab whose weights they hold, Chinese labs' models took about three-fifths of LLM downloads (about 59% to 61%) from February to October 2026, according to our Hugging Face download trends report.
Qwen, Gemma and DeepSeek download statistics from Hugging Face

How Do Qwen, Gemma and DeepSeek Compare Side by Side?

Qwen's share of open LLM downloads was about 4.8 times Gemma's and about 10 times DeepSeek's in the week of Nov. 9, 2026, according to our live LLM download trends. All three are open-weight model families: anyone can download the trained weights, though the training data and code are not necessarily released.

Measure (read Nov. 12, 2026, unless noted) Qwen Gemma DeepSeek
Weekly share, week of Nov. 9, 2026 44.2% 9.3% 4.3%
Weekly share, week of March 16, 2026 48.4% 3.6% 3.5%
Peak weekly share since mid-March 57.7% (weeks of May 4 and 11) 15.6% (week of July 13) 5.9% (week of June 15, during a counting fault); 5.7% outside it
Low weekly share since mid-March 44.2% (week of Nov. 9) 3.5% (week of March 23) 3.2% (week of Sept. 7)
Top model (30 days to Nov. 12, 2026) Qwen3-0.6B, #1, 20.2M gemma-4-31B-it, #7, 6.8M DeepSeek-V3.2, #4, 8.1M
Models in the top 25 16 3 1
Size range in the top 25 0.5B to 35B (3B active) 4B effective to 31B 685B
License of top-25 models Apache-2.0 (all 16) Apache-2.0 (Gemma 4) MIT
Most customized base (six months to Nov. 12, 2026) Qwen3.8-27B gemma-4-26B-A4B-it None in the top 10
Other model types Embeddings, speech, image and video Embeddings (embeddinggemma-2) None

‍

Read the table as a measure of adoption. Qwen spreads its downloads across many models, while DeepSeek concentrates 8.1M on a single 685B model.

Week of Qwen Gemma DeepSeek
March 16, 2026 48.4% 3.6% 3.5%
April 20, 2026 52.5% 8.1% 4.8%
May 11, 2026 57.7% 9.2% 4.2%
July 13, 2026 45.7% 15.6% 5.2%
Sept. 7, 2026 50.5% 10.8% 3.2%
Oct. 5, 2026 48.2% 10.5% 3.9%
Nov. 9, 2026 44.2% 9.3% 4.3%

‍

Since the week of May 11, 2026, the three families' combined share has fallen from 71.1% to 57.8%, and Qwen's has dropped from 57.7% to 44.2%. The gap between Qwen and Gemma was narrowest in the week of July 13, 2026, when Qwen's share was about 2.9 times Gemma's. In the week of Nov. 9, Gemma's share was about 2.2 times DeepSeek's. We computed these ratios from the weekly shares on our live LLM download trends.

Two events distort the series. For about two weeks from mid-April 2026, a burst consistent with a single source added roughly 13 million downloads a day to Qwen3-VL-2B-Instruct, which overlaps Qwen's 57.7% peak. Between June 10 and 19, 2026, almost every model's reported downloads fell by about a third and then recovered, which overlaps DeepSeek's June high. Read both peaks with that in mind.

How Popular Is Qwen in 2026?

Qwen is the most popular open LLM family in 2026 by Hugging Face downloads, according to our live LLM download trends. It leads Gemma and DeepSeek on weekly share, top-25 presence and community customization.

Qwen downloads statistics in 2026

What Share of Open LLM Downloads Does Qwen Hold?

Qwen held 44.2% of open LLM downloads in the week of Nov. 9, 2026, its lowest week since our weekly share series began in mid-March. Its share was 48.4% in the week of March 16 and peaked at 57.7% in the weeks of May 4 and 11, a peak that overlaps the mid-April single-source burst.

Our Hugging Face download trends report shows that Qwen-based models took about 45% to 48% of LLM downloads every month from March to September 2026. The mix inside that share is shifting. Qwen's own uploads fell from about 40% to 31% of LLM downloads, while other uploaders' builds on Qwen doubled from about 8% to 16%. Most of those builds are compressed copies: smaller, lower-precision files of the same weights.

Our data cannot show why Qwen's weekly share has slid to 44.2%.

Which Qwen Models Are Downloaded Most?

Qwen3-0.6B is the most downloaded Qwen model, ranking #1 on our list of the most downloaded open LLMs with 20.2M downloads in the 30 days to Nov. 12, 2026. Over the same 30 days, Qwen held 6 of the top 10 open LLMs and 16 of the top 25. Its top-25 models run from 0.5B to 35B parameters, the widest range of the three families.

Rank Model Size Downloads (30 days to Nov. 12, 2026)
1 Qwen3-0.6B 0.6B 20.2M
3 Qwen3-14B 14B 12.9M
5 Qwen3.8-27B 27B 7.2M
6 Qwen3-VL-8B-Instruct 8B 7.0M
12 Qwen2.5-VL-7B-Instruct 7B 5.9M
19 Qwen2.5-0.5B-Instruct 0.5B 4.4M
21 Qwen3.6-35B-A3B 35B (3B active) 4.1M
23 Qwen3-VL-4B-Instruct 4B 3.6M

‍

The table shows Qwen's top three models, its three vision-language models, which read images as well as text, and its smallest and largest models in the top 25. Four Qwen2.5 models from an earlier generation also remain in the top 25.

Qwen lines also reached the top of 2026's new releases. Of about 2,900 models created in 2026 that reached the top 5,000 by Oct. 1, 46 reached the top 100. Two-thirds of those came from five launch families, according to our report, and three of the five are Qwen lines: Qwen 3.5, 3.6 and 3.8.

How Is Qwen Licensed and Fine-Tuned?

All 16 Qwen models in the top 25 declare Apache-2.0, as read on Nov. 12, 2026.

Customization is where Qwen's lead runs deepest. In the six months to Nov. 12, 2026, 8 of the 10 base models with the most popular custom and compressed versions were Qwen models. Qwen3.8-27B leads with 5 custom versions and 61 compressed copies. Custom versions include fine-tunes, which are trained further on new data.

New community builds keep arriving. Two builds of Qwen 3.8 by autotrust were among the breakout models of Nov. 5 to 12, 2026. Qwen3.8-Flash-Next-E384-FP4-V2 entered the top 5,000 at #3,618 with a trending score of 128, and Qwen3.8-27B-NVFP4-DGX-SPARK-V2 entered at #3,658.

Where Does Qwen Appear Beyond LLMs?

Qwen also appears on three of our other model-type pages:

  • Embeddings and rerankers (updated Nov. 12, 2026): 3.3% of downloads. Qwen3-Embedding-0.6B ranks #11 with 8.9M, Qwen3-Reranker-4B #23 with 2.6M and Qwen3-Embedding-4B #25 with 2.5M.
  • Speech and audio (updated Nov. 12, 2026): 3.8% of downloads, down 3.8 points since Aug. 14.
  • Image and video (updated Nov. 4, 2026): 3.3% of downloads. Qwen-Image-Edit-2509 ranks #17, and three community Qwen-Image builds are also in the top 25.

How Popular Is Gemma in 2026?

Gemma, Google's open model family, is the second most downloaded of the three families, according to our live LLM download trends. It has also gained the most share of the three since mid-March.

What Share of Open LLM Downloads Does Gemma Hold?

Gemma held 9.3% of open LLM downloads in the week of Nov. 9, 2026, up from 3.6% in the week of March 16. Its climb began in the week of April 6 at 5.3% and peaked at 15.6% in the week of July 13. Read the peak and the current share together. Our data cannot show why the share eased after July.

Our report measures the same rise by model owner. Google's monthly share of open LLM downloads went from 3.5% in March 2026 to 13.9% in June and 10.3% in September, and the report attributes the rise to Gemma 4.

Which Gemma Models Are Downloaded Most?

gemma-4-31B-it is the most downloaded Gemma model, at #7 with 6.8M downloads in the 30 days to Nov. 12, 2026. All three Gemma models in the top 25 are Gemma 4, with 16.5M downloads combined.

Rank Model Size Downloads (30 days to Nov. 12, 2026)
7 gemma-4-31B-it 31B 6.8M
15 gemma-4-26B-A4B-it 26B (4B active) 5.4M
20 gemma-4-E4B-it 4B effective (8B total) 4.3M

‍

The "A4B" in gemma-4-26B-A4B-it marks a mixture-of-experts model: it has 26B parameters in total but uses only 4B for each token. Gemma 4 also sits in a growing size class. Four model lines took the 10B to 35B class from about a fifth to about a third of LLM downloads in 2026: Gemma 4 at 26B and 31B, and Qwen 3.6 and 3.8 at 27B, according to our report. Without those four lines, the class is flat.

Gemma 4 is one of the five 2026 launch families behind two-thirds of the year's top-100 newcomers, according to our report.

How Is Gemma Licensed and Fine-Tuned?

Gemma 4 declares Apache-2.0, as read on Nov. 12, 2026. Repos still under the older Gemma license accounted for 1.5% of the most downloaded open LLMs on the same date.

Gemma's customization footprint is small. gemma-4-26B-A4B-it is its only base among the 10 most customized in the six months to Nov. 12, 2026, with 1 custom version and 3 compressed copies.

Where Do Gemma and Google Appear Beyond LLMs?

Gemma shows up in one other category, and Google's other open models are a separate story:

  • Embeddings (updated Nov. 12, 2026): embeddinggemma-2, a new embedding model, drew 492K downloads in 30 days.
  • Translation: Hugging Face tags TranslateGemma as an LLM, so it appears among LLMs rather than on our translation page.
  • Google outside Gemma: 16.2% of vision downloads (updated Nov. 12, 2026, up 5.2 points since Aug. 14), 3.4% of time series forecasting downloads through TimesFM (updated Nov. 9) and the #1 translation and summarization model, t5-small, with 24.8M downloads (updated Oct. 28).

How Popular Is DeepSeek in 2026?

DeepSeek is the third most downloaded of the three families, according to our live LLM download trends. It places fewer models in the top 25 than Qwen or Gemma, but its one entry is the largest model on the list.

What Share of Open LLM Downloads Does DeepSeek Hold?

DeepSeek held 4.3% of open LLM downloads in the weeks of Nov. 2 and Nov. 9, 2026, up from 3.5% in the week of March 16. Its share has stayed between 3.2% and 5.9% every week since. Its 5.9% high, in the week of June 15, falls inside the June 10 to 19 counting fault. Its 5.7% in the weeks of June 8 and June 22 is the safer high.

Its share fell to a 3.2% low in the week of Sept. 7 and has risen since. Our data cannot show why. DeepSeek has held a higher weekly share than Llama, Meta's open model family, every week since the week of Oct. 5, 2026: 3.9% against 3.6% that week, and 4.3% against 2.1% in the week of Nov. 9.

These shares may understate DeepSeek. As our report notes, models that also ship through other channels, such as their own packages, are undercounted in Hugging Face numbers, and private fine-tunes leave no public repo.

Which DeepSeek Models Are Downloaded Most?

DeepSeek-V3.2 is the most downloaded DeepSeek model and its only one in the top 25, ranking #4 with 8.1M downloads in the 30 days to Nov. 12, 2026. That is more than the average top-25 Qwen model (6.6M) or Gemma model (5.5M), both computed from the same window. At 685B parameters, it is the largest model in the top 25.

DeepSeek's newer models are climbing. DeepSeek V4 is one of the five 2026 launch families, according to our report. DeepSeek-V4.1-Flash-1017 climbed 381 places in the top 5,000 between Nov. 5 and Nov. 12, 2026, from #1,401 to #1,020. Its trending score of 164 was the highest among that week's breakout open LLMs.

How Is DeepSeek Licensed and Fine-Tuned?

DeepSeek-V3.2 declares the MIT license, as read on Nov. 12, 2026. No DeepSeek base appears among the 10 most customized in the six months to Nov. 12, 2026, while Qwen and Gemma each have at least one.

Where Does DeepSeek Appear Beyond LLMs?

Across the model-type pages we track, read between Oct. 28 and Nov. 12, 2026, DeepSeek appears only on the LLM page. It has no top-25 model and no maker share in embeddings, speech, vision, text classification, image and video, time series or translation. Teams building on DeepSeek models can work with our DeepSeek developers.

Which Open LLM Should You Choose: Qwen, Gemma or DeepSeek?

Start with the most adopted model for your need, then test alternatives on your own data. Adoption shows where most teams start, not which model fits your data. In the 30 days to Nov. 12, 2026, it points to Qwen for the widest range of sizes and the most fine-tuned versions. It points to gemma-4-31B-it or Qwen3.8-27B for a 26B to 31B model, and to DeepSeek-V3.2 for the largest model or an MIT license.

Need Most-Adopted Starting Point (30 days to Nov. 12, 2026) What to Test Next
Widest size range Qwen, 0.5B to 35B (Qwen3-0.6B at #1 with 20.2M) Another Qwen size, or gemma-4-E4B-it (4B effective)
A 26B to 31B model gemma-4-31B-it (#7, 6.8M) or Qwen3.8-27B (#5, 7.2M) The other of the two, plus gemma-4-26B-A4B-it (#15, 5.4M)
Largest model DeepSeek-V3.2 (#4, 8.1M, 685B) Qwen3.6-35B-A3B, the largest Qwen model in the top 25
Most fine-tuned versions Qwen3.8-27B (5 custom versions, 61 compressed copies) Other Qwen bases among the 10 most customized
MIT license DeepSeek-V3.2 An Apache-2.0 model from Qwen or Gemma 4

‍

Read the table as a measure of adoption. Small models rank high partly because automated test pipelines fetch them, which may help explain Qwen3-0.6B's #1 rank.

Two findings from our report support starting with adoption. Official releases from big labs were about twice as likely as community copies to keep their launch-month downloads three months later. And 70% to 90% of popular derivatives of open LLMs are compressed copies that appear within about a week of a release. Fine-tuned versions follow about two weeks later. Together, these findings suggest that a widely adopted official release tends to hold its downloads and gather derivatives quickly.

Compare two or three candidates on your own data before you switch. Confirm the license on the model card before you deploy or redistribute weights. If you plan to run the winner on your own hardware, our guide to self-hosting an open-weight model covers the trade-offs.

How Did We Measure Open LLM Popularity?

We have recorded Hugging Face's top 5,000 models by 30-day downloads every day since Feb. 12, 2026, with figures refreshed daily at 06:00 UTC. From Feb. 14 to March 12, 2026, we kept only the top 1,000, and June 22 is missing.

A Hugging Face download is a file fetch, usually of a model's config file, not a person. Automated tools and test pipelines count, nothing is deduplicated, and GGUF repos, which package models for running locally, count each file separately. The data covers Hugging Face and public repos only. For these reasons, we compare shares rather than raw totals and state the window for every figure.

Two events distort the series. For about two weeks from mid-April 2026, a burst consistent with a single source inflated Qwen3-VL-2B-Instruct. Between June 10 and 19, 2026, a counting fault cut reported downloads by about a third before they recovered, so we treat that window as estimated.

Our two share measures differ. Our live trends page reports weekly family share, where copies and fine-tunes that name their base model count toward that base's family. Our report credits monthly share to the base-model owner. We keep the two apart in any comparison.

The pages behind this article were updated on different dates: LLMs, embeddings, speech and vision on Nov. 12, 2026; time series on Nov. 9; image and video on Nov. 4; and text classification and translation on Oct. 28. Our Hugging Face download trends report was published Oct. 2, 2026, and covers 1,053,899 rows and 13,182 models from Feb. 12 to Oct. 1, 2026. We recomputed every report finding with independent code, and reviewers then tried to disprove each one.

How Can You Run Qwen, Gemma or DeepSeek on Your Own Data?

You can fine-tune them on your own data. Qwen3-0.6B, gemma-4-31B-it and DeepSeek-V3.2 are general-purpose defaults. They know nothing about your contracts, tickets or product terms, so a model tuned to your own data is the next step.

Valkyrie by Azumo fine-tunes open-weight models on your data and runs them on dedicated hardware that we operate. You buy concurrent agents rather than tokens, so the bill is the same in a heavy month and a quiet one. Valkyrie serves OpenAI- and Anthropic-compatible endpoints, so most tools connect with one environment variable. You own the weights, and the service is SOC 2 certified.

Follow the leaders on our live LLM download trends, updated daily, read the full report, and see how Valkyrie works.

Frequently Asked Questions

  • Most top models in these families declare permissive licenses. All 16 Qwen models in the top 25 use Apache-2.0, Gemma 4 uses Apache-2.0 and DeepSeek-V3.2 uses MIT, as declared on Hugging Face and read on Nov. 12, 2026. Confirm the license on each model card before you deploy or redistribute weights. This is not legal advice.

  • Qwen has the most popular customized versions. In the six months to Nov. 12, 2026, 8 of the 10 base models with the most popular custom and compressed versions were Qwen models, led by Qwen3.8-27B with 5 custom versions and 61 compressed copies. Compressed copies are smaller, lower-precision files with no new training.

  • About three-fifths. Credited to the lab whose weights they hold, Chinese labs' models took about 59% to 61% of LLM downloads from February to October 2026, according to our Hugging Face download trends report. That figure uses share by model owner, which differs from the weekly family share on our live trends page.

Photo of Juan Pablo Lorandi

About the Author:

Chief Technology Officer | Software Architect | Builder of AI, Products, and Teams

Juan Pablo Lorandi is the CTO at Azumo, with 20+ years of experience in software architecture, product development, and engineering leadership.