Gemma 4: Google Open-Sources a 31B That Takes Down 400B, and It Runs on a Mac Mini

Just a couple of days ago, Google DeepMind CEO Demis Hassabis posted a tweet, and the AI world lost its mind.

Gemma 4 is here.

31B parameters, open source, Apache 2.0 license, runs on a Mac Mini, and it beats 400B-parameter giants.

Google had open-sourced Gemma before, but it always fell a bit short — small parameter counts, benchmarks it couldn’t win against competitors, and a pile of licensing headaches. This time, all of that changed.

1. Four Models, One for Everyone

Gemma 4 ships in four variants, running from phones to workstations:

• E2B (2.3B effective parameters / 5.1B including the vocabulary) — phones, Raspberry Pi. Native audio, 128K context. Real-time voice AI on ARM devices is no problem.

• E4B (4.5B effective parameters / 8B including the vocabulary) — laptops, edge devices. Also with audio, 128K context.

• 26B MoE (26B total, 4B activated) — the standout model this time. MoE architecture, activates only 3.8B parameters at inference, so it’s about as fast as a 4B model with quality close to 26B. 256K context. Runs smoothly on a Mac Mini with 24GB of memory.

• 31B Dense (30.7B) — the full-fat workstation flagship. 256K context, no audio but maxed-out visual understanding. Runs bf16 on a single H100, and quantized it runs on consumer GPUs too.

The E in E2B/E4B stands for effective parameters; with the embedding layers added in, inference actually uses 5.1B and 8B. This design relates to Per-Layer Embeddings (PLE, a separate embedding table per layer) and is specifically tuned for on-device memory efficiency.

2. A Generational Leap — How Explosive the Numbers Are

Gemma 4 patches every weakness of the previous generation, and doubles things along the way.

• Math (AIME 2026 no-tools): 20.8% → 89.2%, a gain of 68.4 percentage points. More than a 4x jump. The previous generation basically couldn’t do competition math; Gemma 4 comes close to a perfect score.

• Code (LiveCodeBench v6): 29.1% → 80.0%, up 50.9 points. The previous generation’s coding ability was roughly that of a beginner; Gemma 4 can already serve as half a code assistant.

• Graduate-level scientific reasoning (GPQA Diamond): 42.4% → 84.3%, up 41.9 points. These are PhD-level science questions, and Gemma 4 answers over 80% correctly.

• BigBench Extra Hard: 19.3% → 74.4%, up 55.1 points. Complex tasks requiring multi-step reasoning see an enormous jump.

• Context memory (MRCR v2, 128K long-text retrieval): 13.5% → 66.4%, up 52.9 points. The previous generation’s long context was basically fake — it could accept long input but couldn’t use it well. Gemma 4’s 256K context is genuinely usable.

• Codeforces ELO: 110 → 2150. 110 points is roughly someone who just learned to program; 2150 points is roughly a programmer with competition experience. This kind of leap is extremely rare among open-source models.

Sources: Google’s official model card (ai.google.dev/gemma/docs/core/model_card_4), Stable Learn’s writeup (stable-learn.com/en/gemma-4-model-card/), AI.rs’s head-to-head review (ai.rs/ai-developer/gemma-4-vs-qwen-3-5-vs-llama-4-compared)

3. Arena AI Leaderboard — What Third Parties Say

Google’s own hype doesn’t count, so let’s look at third parties. Arena AI is a crowdsourced chat-preference leaderboard where models don’t know they’re being compared, making it the closest thing to real-world usage rankings.

As of March 31, 2026 (source: Maniac.ai citing the Arena AI page, maniac.ai/blog/qwen-3-5-vs-gemma-4-benchmarks-by-size):

Arena AI is a chat-preference ranking and doesn’t correlate perfectly with static benchmarks. But it does reveal one fact: in real conversational scenarios, Gemma 4’s larger models can already arm-wrestle with a 400B-parameter Qwen.

4. How Does It Compare with Qwen 3.5 — A Head-to-Head at the Same Size

Maniac.ai did a very detailed same-size comparison, taking the benchmarks that both published in their model cards and lining them up row by row. The conclusion is nuanced — it isn’t simply “one crushes the other across the board.”

31B vs 27B (workstation flagship tier)

Qwen is slightly better on static benchmarks, but the gap is tiny; Gemma pulls ahead on multilingual and multimodal. At this size, the two trade blows.

26B MoE vs 35B MoE (consumer MoE tier)

This is Gemma’s strongest showing. The 26B MoE activates only 3.8B parameters at inference, yet reaches 97% of the 31B dense model’s quality:

You pay the inference cost of a 4B model and get 97% of a 31B model’s quality. That efficiency beats any Dense model.

Compared with the equivalent Qwen3.5-35B-A3B (3B activated parameters): on Arena AI, Gemma 26B MoE (1441) leads Qwen 35B MoE (1400) by 41 points — a clear gap.

E2B / E4B (small model tier)

This tier is a bit of a surprise: Qwen 3.5 leads on most benchmarks.

• E4B (4B) vs Qwen3.5-4B: Qwen leads across the board on MMLU-Pro (79.1% vs 69.4%), GPQA (76.2% vs 58.6%), LiveCodeBench (55.8% vs 52.0%), and TAU2 (79.9% vs 42.2%).

• E2B (2B) vs Qwen3.5-2B: same trend, with Qwen ahead on 3 of the 4 overlapping benchmarks.

Gemma E2B/E4B’s advantages are: native audio support (Qwen’s small models don’t have it), the Google mobile ecosystem (Android AICore), and 128K context.

Bottom line: if you’re a small-model person and look purely at text benchmarks, Qwen 3.5 is stronger; if you need audio or want to run on Android, the Gemma E series is your only choice.

5. Can It Run Locally on a Mac?

• 26B MoE (recommended): A Mac Studio M2 Max 64GB handles it with no problem at all; quantized to Q4_K_M it’s about 14GB of VRAM, and a Mac Mini with 24GB of memory can run it too. Speed is roughly 15-25 tokens per second.

• 31B Dense (high performance): On an M2 Max, Q4 quantization takes about 20GB of memory, so you want a 32GB machine for it to be reasonably smooth.

• E4B (everyday laptop use): Any MacBook with 16GB of memory can run it, and 8GB of VRAM is enough. Good for everyday chat, copywriting, and code completion.

Recommended tools: LM Studio (GUI), Ollama (command line), llama.cpp (native inference, best performance).

6. Why This Time Is Different — What the License Means

Gemma 4 switched to Apache 2.0 — what does that mean?

• Use it however you want: any commercial product, any volume, any scenario, completely free.

• Modify it however you want: you can change the code, train on it, distill it, no need to notify Google.

• Distribute it however you want: build a SaaS, run an API service, integrate it into someone else’s product — no need to open-source derivative code.

Compared with Qwen 3.5: also Apache 2.0, so the two are tied.

Compared with Llama 4: the Community License, a 700M MAU cap plus Meta’s Acceptable Use Policy — plenty of restrictions on commercial use. Gemma 4’s license is actually more permissive than Llama 4’s.

For anyone wanting to do indie development or start an AI product company, Gemma 4 basically removes the legal risk.

7. A Few More Technical Details Worth Noting

• Thinking Mode: Gemma 4 supports chain-of-thought reasoning; the model emits its internal reasoning process before answering, up to 4000+ tokens. This follows the technical path of DeepSeek-R1 and OpenAI o1, and is the key to this round’s big gains in math and code.

• Native Function Calling: All variants support structured function calling — no special prompt needed, it returns tool-call instructions directly in JSON. But compared with models that excel at tool calling, like o1 and DeepSeek-R1, Gemma 4’s tool-call success rate still lags — users on GitHub have directly asked Google to strengthen tool-calling ability (on the TAU2 benchmark, Gemma 4 31B scores 76.9%, also below Qwen 3.5’s 79%).

• Per-Layer Embeddings (PLE): Each decoder layer has its own embedding table, one of the key reasons small-parameter models improved.

• Hybrid Attention: The MoE models use a mix of local sliding-window attention and global attention, preserving speed without losing long-context understanding.

• Shared KV Cache: The last few layers share KV tensors, significantly reducing memory usage during long-context inference.

8. How Good Is the Chinese Support — The Question Chinese Users Care About Most

Gemma 4 is a big improvement over the previous generation, but Qwen 3.5 still has the edge in Chinese. Let the data speak.

Gemma 4’s multilingual benchmark (MMMLU) hits 88.4% at 31B, and that progress is real — Gemma 4’s multilingual ability is no longer a weakness.

But AI.rs’s head-to-head review says: “Qwen 3.5 still holds the multilingual crown.” There are two reasons:

• Vocabulary size: Qwen 3.5 uses a 250K vocabulary, specifically optimized for Chinese. Gemma 4 is trained on a mix of 140+ languages, so its vocabulary skews toward general multilingual use, and its Chinese tokenization efficiency isn’t as good as Qwen’s.

• Number of languages: Qwen 3.5 supports 201 languages; Gemma 4 supports 140+. Qwen covers more languages, and especially for Chinese dialects and specific Chinese knowledge, Qwen has both more training data and a higher Chinese share.

Qwen is designed “Chinese-first”; Gemma 4 is designed to treat “all languages equally.” If you mainly handle local Chinese cultural content, internet slang, and Chinese professional knowledge, Qwen 3.5 is still the safer choice.

Is Gemma 4 suitable for Chinese users to deploy?

Can you deploy it? Absolutely. Gemma 4 is a fully open-source Apache 2.0 model, the weights are downloadable directly from HuggingFace, no internet connection required, no Google account required, and no regional restrictions. With Llama you have to worry about US restrictions; Gemma 4 has no such risk.

Is it good? Depends on what you’re doing. If you mainly handle English tasks (writing code, reading English technical docs, English writing), Gemma 4 at the 31B size is extremely strong. If you mainly handle Chinese content, Qwen 3.5 is currently still the better match.

One approach worth considering: a dual-model setup. Use Gemma 4 26B MoE for English code and complex reasoning tasks, and Qwen 3.5 for Chinese conversation and local knowledge.

Gemma 4’s Thinking Mode has full support for Chinese chains of thought — the model can reason step by step in Chinese, which makes for a better experience for users who need to see the reasoning process.

Summary:

• Chinese ability: Gemma 4 is far stronger than the previous generation, and the 88.4% multilingual benchmark is a genuine improvement. But Qwen 3.5’s Chinese vocabulary (250K) and larger Chinese training data mean Chinese scenarios are still Qwen’s home turf.

• Deployment friendliness: Gemma 4 is purely open source (Apache 2.0) with no regional usage restrictions, which is very friendly to Chinese developers.

• Recommended strategy: for pure Chinese scenarios, keep using Qwen 3.5; for English-heavy work or when you need strong reasoning/coding, go with Gemma 4; Mac users should pick Gemma 4 26B MoE — it runs very comfortably locally.

References:

• MMMLU (multilingual general knowledge): Gemma 4 31B 88.4%, data from Google’s official model card (ai.google.dev/gemma/docs/core/model_card_4)

• Qwen multilingual crown assessment: AI.rs, original text “Qwen 3.5 still holds the multilingual crown” — ai.rs/ai-developer/gemma-4-vs-qwen-3-5-vs-llama-4-compared

• Qwen vocabulary data (250K vocab, 201 languages): github.com/QwenLM/Qwen3.5

• Gemma 4 multilingual support (140+ languages): blog.google/innovation-and-ai/technology/developers-tools/gemma-4/

9. References

1. Google official blog — Gemma 4 release announcement

2. Google DeepMind Gemma 4 page — model family overview

3. Google AI Model Card — technical specs and benchmark data

4. AI.rs head-to-head review — Gemma 4 vs Qwen 3.5 vs Llama 4

5. Maniac.ai size-by-size comparison — row-by-row comparison by model size, including third-party Arena AI rankings

6. Stable Learn Gemma 4 writeup — architecture deep dive

7. LM Studio support page — local deployment tool support status

8. HuggingFace download page — 31B instruction-tuned version download

Going from “not quite there” to “genuinely competitive” took Gemma four generations. 400B parameters isn’t the finish line, and 31B isn’t the starting line. The relationship between parameter count and model strength keeps getting broken, over and over again.

— Written April 3, 2026. Sources verified, data has provenance, and uncertain points are noted. Corrections welcome if anything was missed.

Bonus: One Prediction

This moment with Gemma 4 reminds me of an interesting analogy.

The current stage of AI compute development is a bit like the early mainframe era:

1. Client machines couldn’t possibly have high specs

2. As terminals, they had to connect to a remote mainframe and use its compute to do the work

But after a few years of computer development, millions of households could buy a Mac or an IBM machine, all the functionality ran locally, and every service that used to require a network connection moved into a local client.

Mapping that onto today’s AI era, we’re still at the “terminal” stage.

The tokens we buy now all need cloud processing, like paying an electricity bill. But this window won’t last long — ten years at most, and we’ll go from the “terminal” era to the “home computer” era, where tokens all run locally.

The emergence of open-source models like Gemma 4 is accelerating that process.