Model comparison
Gemma 4 12B vs Qwen3.5 397B
Head-to-head evidence from 17 shared benchmark results across 7 categories. Overall scores shown here use the public BenchAlign v5 ranking lane.
Public leaderboard positions: Gemma 4 12B #137 (Estimated); Qwen3.5 397B #71 (Estimated). Intervals and evidence labels describe ranking uncertainty, not a guarantee for a specific workload.
Evidence parity. Gemma 4 12B and Qwen3.5 397B share 17 comparable benchmark results. 4 of 8 categories are comparable. 6 results are unique to Gemma 4 12B; 38 to Qwen3.5 397B.
Updated July 23, 2026- Shared results
- 17
- Gemma 4 12B only
- 6
- Qwen3.5 397B only
- 38
- Comparable categories
- 4 / 8
Pick Qwen3.5 397B if you want the stronger benchmark profile. Gemma 4 12B only becomes the better choice if knowledge is the priority or you need the larger 256K context window.
Confidence note. This is a partial-evidence comparison with 17 shared benchmark results across 7 evidence categories; 4 of 8 categories currently have scoreable aggregates for both models. Treat the verdict as directional until coverage is more balanced.
Why this result
Qwen3.5 397B is clearly ahead on the BenchAlign aggregate, 57.01 to 47.29. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
Qwen3.5 397B's sharpest advantage is in reasoning, where it averages 63.2 against 43.4. The single biggest benchmark swing on the page is AIME26, 77.5% to 93.3%. Gemma 4 12B does hit back in knowledge, so the answer changes if that is the part of the workload you care about most.
Gemma 4 12B is the reasoning model in the pair, while Qwen3.5 397B is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. Gemma 4 12B gives you the larger context window at 256K, compared with 128K for Qwen3.5 397B.
Category breakdown
Exact category averages are shown below. Not measured means BenchLM does not have enough sourced public coverage for that model and category.
| Category | Gemma 4 12B | Δ | Qwen3.5 397B |
|---|---|---|---|
| Knowledge | Gemma 4 12B77.5 | Margin← 20.9 | Qwen3.5 397B56.6 |
| Reasoning | Gemma 4 12B43.4 | Margin→ 19.8 | Qwen3.5 397B63.2 |
| Math | Gemma 4 12B77.5 | Margin→ 13.1 | Qwen3.5 397B90.6 |
| Multimodal | Gemma 4 12B69.1 | Margin→ 10.5 | Qwen3.5 397B79.6 |
| Agentic | Gemma 4 12BNot measured | MarginNo overlap | Qwen3.5 397B56.5 |
| Coding | Gemma 4 12BNot measured | MarginNo overlap | Qwen3.5 397B66.5 |
| Multilingual | Gemma 4 12BNot measured | MarginNo overlap | Qwen3.5 397B84.7 |
| Inst. Following | Gemma 4 12BNot measured | MarginNo overlap | Qwen3.5 397B92.6 |
Decisive benchmark drivers
The largest measured benchmark gaps in this matchup, with exact reported values.
More
- Source ↗
AIME26
MathA 77.5%B 93.3%Winner: Qwen3.5 397BΔ 15.8AIME26: Gemma 4 12B scored 77.5%; Qwen3.5 397B scored 93.3%. Qwen3.5 397B wins this benchmark. - Source ↗
MMLU-Pro
KnowledgeA 77.2%B 87.8%Winner: Qwen3.5 397BΔ 10.6MMLU-Pro: Gemma 4 12B scored 77.2%; Qwen3.5 397B scored 87.8%. Qwen3.5 397B wins this benchmark. - Source ↗
MMMU-Pro
MultimodalA 69.1%B 79%Winner: Qwen3.5 397BΔ 9.9MMMU-Pro: Gemma 4 12B scored 69.1%; Qwen3.5 397B scored 79%. Qwen3.5 397B wins this benchmark. - Source ↗
GPQA
KnowledgeA 78.8%B 88.4%Winner: Qwen3.5 397BΔ 9.6GPQA: Gemma 4 12B scored 78.8%; Qwen3.5 397B scored 88.4%. Qwen3.5 397B wins this benchmark.
Operational comparison
Runtime and commercial metrics are compared only when both models have a complete sourced value.
| Metric | Gemma 4 12B | Qwen3.5 397B | Comparison |
|---|---|---|---|
| Input / output priceUSD per 1M tokens | Gemma 4 12BNot available | Qwen3.5 397B$0.6 input / $3.6 output | A complete price comparison is not available. |
| Generation speedtokens per second | Gemma 4 12BNot available | Qwen3.5 397B96 tok/s | A complete speed comparison is not available. |
| First-answer latencyseconds to first token | Gemma 4 12BNot available | Qwen3.5 397B2.44 s | A complete latency comparison is not available. |
| Context windowmaximum listed tokens | Gemma 4 12B256K | Qwen3.5 397B128K | Gemma 4 12B lists the larger context window. |
Benchmark Deep Dive
Agentic18 benchmarks
| Benchmark | Gemma 4 12B | Qwen3.5 397B | Result |
|---|---|---|---|
| τ²-bench resultsSource | 36.3% | 95.6% | Qwen3.5 397B leads |
| Terminal-Bench 2.0Source | — | 52.5% | Not comparable |
| BrowseCompSource | — | 62% | Not comparable |
| Claw-EvalSource | — | 56.8% | Not comparable |
| QwenClawBenchSource | — | 51.8% | Not comparable |
| τ³-bench resultsSource | — | 68.4% | Not comparable |
| VITA-BenchSource | — | 43.7% | Not comparable |
| DeepPlanningSource | — | 37.6% | Not comparable |
| ToolathlonSource | — | 36.3% | Not comparable |
| MCP AtlasSource | — | 46.1% | Not comparable |
| MCP-TasksSource | — | 74.2% | Not comparable |
| WideResearchSource | — | 74.0% | Not comparable |
| Gert LabsSource | — | 46.76% | Not comparable |
| ResearchClawBenchSource | — | 14.2% | Not comparable |
| AA Agentic IndexSource | — | 19.9% | Not comparable |
| APEX-Agents-AASource | — | 15.3% | Not comparable |
| GDPval-AASource | — | 23.1% | Not comparable |
| GDPval-AASource | — | 962 | Not comparable |
Coding5 benchmarks
ReasoningQwen3.5 397B wins6 benchmarks
KnowledgeGemma 4 12B wins15 benchmarks
| Benchmark | Gemma 4 12B | Qwen3.5 397B | Result |
|---|---|---|---|
| GPQASource | 78.8% | 88.4% | Qwen3.5 397B leads |
| GPQA-DSource | 78.8% | — | Not comparable |
| MMLU-ProSource | 77.2% | 87.8% | Qwen3.5 397B leads |
| HLE w/o toolsSource | 5.2% | — | Not comparable |
| MMMLUSource | 83.4% | — | Not comparable |
| Artificial Analysis Intelligence IndexSource | 22.0% | 33.7% | Qwen3.5 397B leads |
| AA-GPQA DiamondSource | 75.3% | 89.3% | Qwen3.5 397B leads |
| AA-HLESource | 14.8% | 27.3% | Qwen3.5 397B leads |
| AA-Omniscience IndexSource | -51.9% | -29.8% | Qwen3.5 397B leads |
| AA-Omniscience AccuracySource | 16.0% | 31.4% | Qwen3.5 397B leads |
| AA-Omniscience Hallucination RateSource | 80.8% | 89.1% | Gemma 4 12B leads |
| SuperGPQASource | — | 70.4% | Not comparable |
| MMLU-ReduxSource | — | 94.9% | Not comparable |
| C-EvalSource | — | 93% | Not comparable |
| HLESource | — | 28.7% | Not comparable |
MathQwen3.5 397B wins5 benchmarks
Multilingual2 benchmarks
MultimodalQwen3.5 397B wins8 benchmarks
| Benchmark | Gemma 4 12B | Qwen3.5 397B | Result |
|---|---|---|---|
| MMMU-ProSource | 69.1% | 79% | Qwen3.5 397B leads |
| MathVisionSource | 79.7% | 88.6% | Qwen3.5 397B leads |
| MedXpertQA (MM)Source | 48.7% | — | Not comparable |
| AA-MMMU-ProSource | 69.7% | 77.3% | Qwen3.5 397B leads |
| CharXivSource | — | 80.8% | Not comparable |
| VideoMMMUSource | — | 84.7% | Not comparable |
| ScreenSpot ProSource | — | 65.6% | Not comparable |
| V*Source | — | 95.8% | Not comparable |
Frequently Asked Questions (5)
Which is better, Gemma 4 12B or Qwen3.5 397B?
Qwen3.5 397B is ahead on BenchLM's BenchAlign leaderboard, 57.01 to 47.29. The biggest single separator in this matchup is AIME26, where the scores are 77.5% and 93.3%.
Which is better for knowledge tasks, Gemma 4 12B or Qwen3.5 397B?
Gemma 4 12B has the edge for knowledge tasks in this comparison, averaging 77.5 versus 56.6. Inside this category, AA-Omniscience Index is the benchmark that creates the most daylight between them.
Which is better for math, Gemma 4 12B or Qwen3.5 397B?
Qwen3.5 397B has the edge for math in this comparison, averaging 90.6 versus 77.5. Inside this category, AIME26 is the benchmark that creates the most daylight between them.
Which is better for reasoning, Gemma 4 12B or Qwen3.5 397B?
Qwen3.5 397B has the edge for reasoning in this comparison, averaging 63.2 versus 43.4. Inside this category, AA-LCR is the benchmark that creates the most daylight between them.
Which is better for multimodal and grounded tasks, Gemma 4 12B or Qwen3.5 397B?
Qwen3.5 397B has the edge for multimodal and grounded tasks in this comparison, averaging 79.6 versus 69.1. Inside this category, MMMU-Pro is the benchmark that creates the most daylight between them.
Related Comparisons
Explore More
Choose a model with this week’s evidence
Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.
One email each week. Unsubscribe anytime.