# Gemma vs Qwen: Gemma 4 vs Qwen 3.5, 3.6 and 3.8, Side by Side

> Google's and Alibaba's own published numbers for Gemma 4 and Qwen 3.5, 3.6 and 3.8: coding scores, the sizes you can run at home, languages, audio and licenses.

By Pat Simmons · November 9, 2026 · 8 min read

Canonical: https://www.aiformortals.co/blog/gemma-vs-qwen

Pick Qwen if coding or hard reasoning is the job: in Qwen's own model card, Qwen3.6-35B-A3B scores 73.4 on SWE-bench Verified against 52.0 for Gemma 4 31B. Pick Gemma 4 if you want a very small model or one that listens to audio, since its E2B and E4B sizes take spoken audio and Google designed them for laptops and phones. Both are open-weight models under Apache 2.0, which means the model files are free to download and run yourself; every score on this page comes from Google or Alibaba, and AI for Mortals hasn't run either model.

Last checked November 9, 2026: the newest releases on Qwen's Hugging Face page (Qwen 3.8, out since August 2026) and in Google's Gemma release notes (Gemma 4, whose 12B size arrived on June 3, 2026), plus the scores, sizes and licenses on both companies' model cards.

## Gemma vs Qwen: which to pick, by job

On the companies' published numbers, Qwen leads for coding, the two are close on knowledge and reasoning, and Gemma 4 is the choice for small devices and audio.

Writing and fixing code: Qwen. The biggest gap in Qwen's table is on SWE-bench Verified, a test built from real bug reports in public code projects, where Qwen3.6-35B-A3B scores 73.4, Gemma 4 31B 52.0 and Gemma 4 26B A4B 17.4.

Science and math questions: close, with Qwen slightly ahead in its own table. Qwen3.6-35B-A3B scores 86.0 on GPQA Diamond (graduate-level science questions) to Gemma 4 31B's 84.3, and 92.7 on AIME26 (a US math competition) to 89.2.

General knowledge: a tie. Both Qwen3.6-35B-A3B and Gemma 4 31B score 85.2 on MMLU-Pro, a broad multiple-choice exam.

Phones, small laptops and audio: Gemma 4. Google built the E2B and E4B sizes for on-device use, publishes how much memory each one needs (see the sizes section below), and gives E2B, E4B and 12B native audio input. None of the three Qwen model cards checked for this page mentions audio input.

Very long documents: Qwen, on paper. Its cards list 262,144 tokens (the chunks of text a model reads) natively and say that extends to about 1 million, while Gemma 4 tops out at 256K.

## Gemma vs Qwen for coding

Qwen leads on coding in Qwen's own published comparison, by a wide margin on fixing real code and by under a point on competitive programming puzzles. Google's Gemma 4 card doesn't report SWE-bench at all, so the biggest gap below rests only on Qwen's measurement.

The table is copied from Qwen's Qwen3.6-35B-A3B model card. For the four tests Google also publishes, Qwen's Gemma figures match Google's Gemma 4 model card exactly.

Test (reported by Qwen) Qwen3.6-35B-A3B Gemma 4 31B Gemma 4 26B A4B
SWE-bench Verified (fixing real code)73.452.017.4
LiveCodeBench v6 (coding puzzles)80.480.077.1
GPQA Diamond (science)86.084.382.3
AIME26 (math)92.789.288.3
MMLU-Pro (general knowledge)85.285.282.6

Qwen's notes say it ran the SWE-bench tests with its own internal agent setup, which gives the model a command line and file-editing tools. Google's card gives its own coding numbers instead: 80.0 on LiveCodeBench v6 for Gemma 4 31B and a Codeforces rating of 2150, a score from a competitive programming site.

_Qwen's own chart from its Qwen3.6-35B-A3B model card: Qwen models carry Qwen's logo, Gemma 4 models carry Google's. Source: Qwen on Hugging Face._

## Qwen 3.5 vs Gemma 4: what's changed since

Qwen 3.5 came out in February 2026 and Gemma 4 on March 31, and Qwen has since released Qwen 3.6 in April and Qwen 3.8 in August. Google has since added a 12B size to Gemma 4, plus MTP (multi-token prediction) versions built to answer faster.

The Qwen 3.5 vs Gemma 4 matchup still has a head-to-head table, because Qwen kept a Qwen3.5-27B column in the Qwen 3.6 card. In that table Qwen3.5-27B beats Gemma 4 31B on SWE-bench Verified (75.0 to 52.0), GPQA (85.5 to 84.3) and AIME26 (92.6 to 89.2). Gemma 4 31B comes out ahead on Terminal-Bench 2.0 (42.9 to 41.6) and on QwenWebBench, Qwen's own test of building web pages (1197 to 1068).

Google's launch chart uses a different measure. It plots scores from Arena.ai's chat arena, a public leaderboard, against model size, and puts Gemma 4 31B above every Qwen 3.5 model shown, including the far larger Qwen3.5-397B-A17B. The same launch post says the 31B model ranked third among open models there and the 26B model sixth.

_Google's chart from the Gemma 4 launch post: Arena.ai chat scores against model size, with Gemma 4 31B and 26B A4B highlighted and three Qwen 3.5 models plotted. Source: Google._

For Qwen 3.8 there's no Gemma comparison to show. The Qwen3.8-27B model card compares it with Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B and Opus 4.6 Max, and reports 89.2 on GPQA Diamond and 90.3 on LiveCodeBench v6. Neither company has published a table with Qwen 3.8 and Gemma 4 side by side, and scores from two different tables can come from different test setups.

## Which sizes can run on your own computer?

Google publishes the memory each Gemma 4 size needs; Qwen's model cards don't. Google's figures cover loading the model only, and longer conversations need more on top.

Quantized means a compressed copy of the model that uses less memory in exchange for a small loss in precision. The figures below are from Google's Gemma 4 overview.

Gemma 4 sizeFull 16-bit8-bit4-bitMobile version
E2B11.4 GB5.7 GB2.9 GB1.1 GB
E4B17.9 GB8.9 GB4.5 GB2.5 GB
12B26.7 GB13.4 GB6.7 GBNot listed
26B A4B57.7 GB28.8 GB14.4 GBNot listed
31B69.9 GB34.9 GB17.5 GBNot listed

The 26B A4B is a mixture-of-experts model, which uses only about 4 billion of its parameters for each token it writes. Google notes that all 26 billion still have to be loaded into memory, so its memory needs are much closer to a regular 26B model than a 4B one.

Google's launch post says the E2B and E4B models run fully offline on phones, the Raspberry Pi and NVIDIA's Jetson Orin Nano, and that quantized 26B and 31B models run on consumer graphics cards. Google's docs list LM Studio, Ollama, llama.cpp and MLX as ways to run Gemma locally.

On the Qwen side, Hugging Face listed 1,729 quantized versions of Qwen3.8-27B as of November 9, 2026, and offers them for apps such as llama.cpp, Ollama and LM Studio. Qwen's cards don't say how much memory any of them needs.

## Languages, images, audio and context length

Qwen lists more languages and a longer stretch of text per conversation; Gemma 4 is the one that takes audio. Both read images, and both read video on at least some sizes.

 Gemma 4 Qwen 3.5 to 3.8
Newest releaseMarch 31, 2026 (12B size June 3, 2026)Qwen 3.8, August 2026
SizesE2B, E4B, 12B, 26B A4B, 31BQwen 3.8 in 27B, Flash-Next and 2.4T-A95B; Qwen 3.6 in 27B and 35B-A3B
Context window128K on E2B and E4B, 256K on the rest262,144 tokens, extendable to about 1 million
InputsText and images on every size; audio on E2B, E4B and 12BText, images and video
LanguagesTrained on 140+, 35+ supported out of the box201 languages and dialects (Qwen 3.5)
LicenseApache 2.0Apache 2.0

The context window is how much text the model can hold in one conversation, counted in tokens. Gemma 4's 256K and Qwen's 262,144 are the same size written two ways, so for the larger models the difference is Qwen's extension to about 1 million. The language count for Qwen comes from the Qwen3.5-27B card; the Qwen 3.6 and 3.8 cards don't give one.

Google's launch post says every Gemma 4 size handles video, and the Qwen3.8-27B card says it understands video up to hour-long clips. Qwen 3.8 thinks step by step before answering by default, and its card shows how to switch that off for a single request; Google says every Gemma 4 model has thinking modes you can configure.

For each new release as it lands, see our Gemma updates page and Qwen updates page.

## Licenses and what they cost

Both are free to download under the Apache 2.0 license, which allows commercial use, so the cost is the hardware you run them on or a hosting service's bill. Google's model card and its Gemma 4 31B repository on Hugging Face both list Apache 2.0, and Google's launch post calls it a commercially permissive open-source license. Google's docs also link a separate prohibited-use policy for Gemma.

Every Qwen card checked for this page (3.5, 3.6 and 3.8) lists Apache 2.0. The Qwen3.8-27B card says a hosted version on Qwen Cloud is coming soon, and Google offers Gemma 4 through Vertex AI. This page doesn't list hosted prices for either.

## Where these numbers come from

Every score on this page is published by the company that makes the model, and AI for Mortals hasn't tested Gemma or Qwen. Google's numbers come from its Gemma 4 model card and launch post, and Qwen's come from its Hugging Face model cards, including the table where Qwen scored Gemma 4.

Companies choose the tests and the settings for their own tables. Qwen's notes on the Qwen 3.6 card list its setups, for example that AIME26 uses the full 2026 exam (parts I and II), so its scores may differ from those in the Qwen 3.5 notes.

_Google's own Gemma 4 scores from the launch post, compared with Gemma 3 27B. The LiveCodeBench, GPQA Diamond and AIME 2026 figures are the same ones Qwen quotes for Gemma. Source: Google._

Pat runs his own head-to-head builds on some open-weight models, and his Kimi K3.1 Pro review shows how Kimi did on the same jobs as Opus 5.5, GPT-6 Astra and others.

## Questions people ask

Is Gemma 4 free?

Yes. The model files are free to download from Hugging Face or Kaggle under Apache 2.0, and you pay only for the computer or hosting you run them on.

Is Qwen free?

Yes. Qwen 3.5, 3.6 and 3.8 are all published on Hugging Face under Apache 2.0.

Gemma 4 vs Qwen 3.6: which scores higher?

Qwen3.6-35B-A3B, in Qwen's own table: it leads Gemma 4 31B on coding, science and math tests and ties it on MMLU-Pro. The coding section above has the numbers.

Gemma 4 vs Qwen 3.8: has anyone compared them?

Neither Google nor Qwen has published one. Qwen's 3.8 card compares it only with other models, so any Gemma 4 vs Qwen 3.8 call today would mix numbers from different tables.

## Sources

- Qwen: Qwen3.6-35B-A3B model card, with Gemma 4 comparison (checked November 9, 2026)

- Qwen: Qwen3.8-27B model card (checked November 9, 2026)

- Qwen: Qwen3.5-27B model card (checked November 9, 2026)

- Qwen on Hugging Face: model list (checked November 9, 2026)

- Google: Gemma 4 model card (checked November 9, 2026)

- Google: Gemma 4 overview and memory requirements (checked November 9, 2026)

- Google: Gemma release notes (checked November 9, 2026)

- Google: Gemma 4 launch post

- Google: Gemma 4 31B on Hugging Face (license)