Skip to content
Andrew Voirol
WorkLogAboutContact
HomeWorkLog

AboutContact

✦ Just one prompt away from figuring it all out.

Work → Gemma 4 Benchmark Suite
ResearchLive

Gemma 4 Benchmark Suite

450+ tests across 7 models on two machines — iMac 2017 (Vulkan) and MacBook M4 Pro (Metal). GPU-accelerated to 90 tok/s.

450+ tests · 7 models · 94% top score

Started Apr 6, 2026·Updated Jun 2, 2026
View Source

The Gemma 4 Benchmark Suite is a comprehensive testing pipeline designed to answer a single question: what does a state-of-the-art local AI model actually feel like on consumer hardware? Instead of relying on cloud data centers or enterprise GPUs, this suite ran over 450 tests across seven different models on two machines — an 8-year-old 2017 iMac and a modern MacBook M4 Pro.

The results challenged conventional wisdom. By utilizing the LunarG Vulkan SDK to resurrect a long-forgotten Radeon Pro 575 GPU, the iMac reached conversational speeds of 37.6 tok/s. On the Apple Silicon side, the M4 Pro ripped through inference at a blazing 90 tok/s for smaller models. More importantly, the largest 31B model—widely deemed impossible to run on older hardware—achieved a remarkable 94% top score in complex reasoning tasks, running completely unsupervised overnight.

The methodology prioritized real-world capability over purely synthetic numbers. Testing covered everything from logic and coding to multi-turn coherence and tool calling, with results processed through deterministic evaluation and LLM-as-a-judge pipelines. All runs were fully autonomous, orchestrated via bash scripts that managed execution, validation, and JSON reporting without human intervention.

Key Results

ModelScoreSpeedVRAM
Gemma 4 31B (E4B)94%7.5 tok/s (iMac Vulkan)4 GB
Gemma 4 12B (E2B)85%38 tok/s (iMac) / 90 tok/s (M4 Pro)2 GB
Gemma 4 26B-A4B (MoE)—Blocked by Vulkan MoE bug4 GB
94%

Visual Analysis

The Data, Visualized

Seven models tested on the same iMac. Same tests, same grading. The story is in the shapes.

0

test results

across 7 models

0

universally perfect

out of 0 tests

0

tok/s peak

E2B Q4

0

min longest test

⚡

Daily Driver

E2B Q4

Fastest + highest score

92%6.95 tok/s · 7.2 GB
💎

Best Value

Unsloth DQ4

17.9% per GB

91%3.41 tok/s · 5.1 GB
👑

Quality King

31B Q4

93% accuracy

93%0.6 tok/s · 19.9 GB

Model Fingerprints

Each model has a shape. Click to compare. Hover to isolate.

PerformanceReasoningCodingTool Calling

Speed vs Accuracy

Bubble size = disk footprint. Top-right is the sweet spot.

89%91%93%1357tokens/sec →accuracy →↗ fast + smartE2B Q4E2B Q8E2B F16E4B Q8E4B F16DQ431B Q4

The Runtime Race

Same 39 tests. The fastest model finishes before the slowest has loaded.

E2B Q8
36 min
92%
E2B Q4
38 min
92%
Unsloth DQ4
76 min
91%
E4B Q8
85 min
90%
E2B F16
149 min
91%
E4B F16
295 min
90%
31B Q4
13h 46m
93%

Wall-clock time for 39 identical tests · score at right

The Inverted Ladder

Less precision. Higher score. The counterintuitive finding.

Q4
92.2%
6.95 tok/s↑
7.2 GB · 38 min
Q8
91.8%
6.95 tok/s
8.1 GB · 36 min
F16
91.2%
1.72 tok/s
10.3 GB · 149 min

Same 5.1B architecture. Same 39 tests. Same iMac.
Lower precision → higher score.


Spotlight

Reasoning Tasks — The Heavy Lifts

The longest and hardest reasoning tasks across all models. Think mode on, math hard.

gemma4:31bAIME Math: Number Theory16.3m100%
gemma4:31bLogic: Constraint Satisfaction31.0m100%
gemma4:31bMulti-Tool: Research + Email15.0m100%
gemma4:31bContext Stress: Needle in Haystack11.4m100%
gemma4:31beBay Buyer Negotiation15.6m100%
gemma4:31bChart Reading: Bar Chart15.1m70%

Deep Dive

Experiments beyond the scores

Beyond the standard benchmarks, we ran targeted experiments to answer specific questions about how these models behave in practice. Expand each to dig in.

We ran the same tasks with Think Mode ON and OFF across E2B and 26B models. The results were surprising — thinking doesn't always help, and sometimes it actively hurts.

TaskE2B OffE2B Think26B Off26B Think
Math100%20%100%100%
Logic20%20%60%20%
Code70%70%70%70%
Creative100%100%100%85%

Think mode hurt E2B on math (100% → 20%) and 26B on logic (60% → 20%). On this hardware, the extra tokens spent "thinking" can actually degrade quality.

We buried a passphrase in progressively larger documents to find the real context ceiling on this hardware. The advertised limits don't match reality.

1K tokens

E2B

✓

26B

✓

4K tokens

E2B

✓

26B

✗

8K tokens

E2B

✗

26B

✗

16K tokens

E2B

✗

26B

✗

On 40GB RAM, E2B tops out around 4K tokens reliably. The 26B can only handle ~1K before OOM pressure causes timeouts. Forget about the advertised 128K context window on consumer hardware.

Can the models maintain coherence across a 5-turn conversation? Both E2B and 26B scored 95% — dropping only one point on turn 4 (a follow-up question that required referencing context from turn 1).

E2B
95%
26B
95%

Both models: turn scores [100, 100, 100, 75, 100]. Impressive coherence.


Full Test Suite

Every test, every model, every score

All 39 unique tests across 7 model configurations. Click column headers to sort. Filter by category to focus on what matters to you.

TestCatAvg ↓

0 tests shown · 0 total results


All benchmarks were run locally using Ollama on a 2017 iMac (i7-7700K, 40GB RAM, macOS). Nothing fancy — this is a machine I actually use. Each test was executed sequentially with cold-start measurements, no prompt caching, and real wall-clock timing.

Models tested span Google's Gemma 4 family from the smallest E2B (2 billion parameters) up to the full 31B. Tests cover 10 categories: performance profiling, mathematical reasoning (AIME-style problems), code generation, tool calling, creative writing, multimodal analysis, agentic task completion, parameter optimization, context window stress testing, and score calibration.

Scoring uses a 0-1 scale. Performance tests are scored on metrics only (latency, throughput). Reasoning and coding tests use automated verification against expected answers.

I'm not a research lab. I'm one person who wanted to know if these models are actually useful on hardware I already own. I publish the failures alongside the wins because that's what's actually helpful.

— The whole point

● 7 models testedSome tests are still running. Some results will change as I learn more. The messy build log is more useful than the polished launch post.

Full source: prompts, grading, raw results→

For the complete story behind these numbers, including the MoE bug hunt and the three-way quantization showdown, read the full Gemma 4 Benchmarks Thread.

AIBenchmarksGemma 4Local-firstVulkanApple Silicon

Related Threads

Running Gemma 4 on Consumer Hardware

The full story of benchmarking Gemma 4 across a 2017 iMac and MacBook M4 Pro — GPU acceleration, quantization showdowns, overnight autonomous pipelines, and the model everyone said couldn't run scoring 94%.


Andrew Voirol

Builder, hacker, shipper. Currently leaving localhost.

Navigate

WorkBuilder’s LogAboutContactRSS Feed

Connect

X / TwitterGitHubLinkedIn

© 2026 Andrew Voirol✦Just one prompt away from figuring it all out.