Compare / Gemma 3 12B vs Gemma 3 4B
Gemma 3 12B vs Gemma 3 4B
Which is cheaper, which is more capable, and which does better at each kind of work, from 29 results on 5 sources that measured both models.
Of the 29 results both models have, the better value on each. Results whose reported ranges overlap are too close to call.
Results as of 16 November 2026 (catalogue release 2026-11-16-49bdfa045dee).
Gemma 3 12B or Gemma 3 4B? The short answer
They score about the same, so pick Gemma 3 4B: it costs 43% less per token.
- Gemma 3 4B is cheaper: $0.042 against $0.075 per million tokens, 43% less.
- Gemma 3 4B scores higher on the Artificial Analysis Intelligence Index: 4.8 against 3.8.
- On third-party benchmarks, on the results both models have: Gemma 3 12B does better for agents and tool use, writing and sticking to the facts; Gemma 3 4B for reasoning and knowledge.
Head to head
Every fact side by side
| Gemma 3 12B | Gemma 3 4B | |
|---|---|---|
| Price per million tokens | $0.050 in · $0.15 out | $0.040 in · $0.050 out |
| Blended (3 in : 1 out) | $0.075 | $0.042 |
| Cheapest host, blended | $0.075 (DeepInfra, BF16) | $0.042 (Parasail, BF16) |
| Intelligence Index | 3.8 | 4.8 |
| Spring Prompt overall | – | – |
| Output speed, tokens a second | – | – |
| Context window | 131,072 tokens | 131,072 tokens |
| Released | 11 Mar 2025 | 11 Mar 2025 |
| Weights | Open | Open |
| Developer based in | the US | the US |
Prices are the developer's list price where we have read it, otherwise the typical host on OpenRouter; see every model's price. Intelligence and speed from Artificial Analysis.
Which is better for what
Each model's rank among the models on sale today, for each kind of work, from our own benchmarks and licensed sources. Further right is better; the better rank is in green.
Every shared result
Each model's best result on every measure both have, grouped by benchmark, with the setting that produced it. Under each measure, every model that benchmark has measured is a grey tick, best to the right, so you can see where the two sit in the field. The headline measures of each benchmark come first; the rest are under "Show all". Green marks the better value; where a source reports ranges that overlap, neither is marked.
| Measure | Gemma 3 12B | Gemma 3 4B |
|---|---|---|
| Arena (formerly LMArena) reported by Arena (formerly LMArena) · 3 results | ||
| Overall · TextArena rating, higher is better 204 · 255 of 387 measured | 1,342 | 1,303 |
| Business, management and finance · TextArena rating, higher is better 155 · 208 of 380 measured | 1,386 | 1,330 |
| Creative writing · TextArena rating, higher is better 173 · 243 of 385 measured | 1,333 | 1,275 |
| Artificial Analysis reported by Artificial Analysis · 3 results | ||
| Artificial Analysis Intelligence Indexindex score, higher is better 249 · 246 of 249 measured | 3.8 | 4.8 |
| GPQA Diamond% of questions, higher is better 225 · 227 of 232 measured | 34.9% | 29.1% |
| Humanity's Last Exam% of questions, higher is better 221 · 194 of 250 measured | 4.2% | 5.3% |
| Berkeley Function Calling Leaderboard (BFCL) V4 reported by Berkeley Function Calling Leaderboard (BFCL) V4 · 3 results | ||
| Overall accuracy · BFCL% correct, higher is better 48 · 74 of 81 measured | 30.4%bfcl prompt | 19.6%bfcl prompt |
| Irrelevance detection · BFCL% correct, higher is better 62 · 71 of 81 measured | 70.3%bfcl prompt | 53.9%bfcl prompt |
| Memory · BFCL% correct, higher is better 18 · 53 of 81 measured | 27.5%bfcl prompt | 8.6%bfcl prompt |
| UGI Leaderboard reported by UGI Leaderboard · 3 results | ||
| Writing score · UGIscore out of 100, higher is better 156 · 185 of 230 measured | 29.9 | 24.0 |
| Requested-length error · UGI% off the requested word count, lower is better 33 · 33 of 230 measured | 9.0% | 9.0% |
| Style adherence · UGIscore from 0 to 1, higher is better 212 · 214 of 230 measured | 0.29 | 0.29 |
| Vectara Hallucination Leaderboard reported by Vectara Hallucination Leaderboard · 2 results | ||
| Hallucination rate · Vectara% of summaries, lower is better 7 · 28 of 104 measured | 4.4% | 6.4% |
| Answer rate · Vectara% of documents, higher is better 82 · 103 of 104 measured | 97.4% | 67.3% |
Show all 29 results (15 more)
| Measure | Gemma 3 12B | Gemma 3 4B |
|---|---|---|
| Arena (formerly LMArena) reported by Arena (formerly LMArena) · 3 results | ||
| Expert prompts · TextArena rating, higher is better 257 · 270 of 338 measured | 1,274 | 1,260 |
| Instruction following · TextArena rating, higher is better 207 · 270 of 387 measured | 1,321 | 1,267 |
| Writing, literature and language · TextArena rating, higher is better 208 · 237 of 386 measured | 1,313 | 1,298 |
| Artificial Analysis reported by Artificial Analysis · 7 results | ||
| Terminal-Bench 2.1% of tasks, higher is better 130 · 127 of 131 measured | 0.0% | 0.4% |
| τ-bench banking% of tasks, higher is better 119 · 121 of 121 measured | 0.8% | 0.4% |
| Artificial Analysis Coding Indexindex score, higher is better 125 · 129 of 130 measured | 5.8 | 2.7 |
| IFBench% of instructions, higher is better 151 · 187 of 195 measured | 36.7% | 28.3% |
| Long-context reasoning (AA-LCR)% of questions, higher is better 224 · 226 of 241 measured | 8.3% | 6.7% |
| Terminal-Bench Hard% of tasks, higher is better 180 · 180 of 190 measured | 0.8% | 0.8% |
| τ²-bench telecom% of tasks, higher is better 183 · 187 of 193 measured | 10.8% | 5.0% |
| Berkeley Function Calling Leaderboard (BFCL) V4 reported by Berkeley Function Calling Leaderboard (BFCL) V4 · 5 results | ||
| Multi-turn tasks · BFCL% correct, higher is better 60 · 75 of 81 measured | 5.8%bfcl prompt | 0.4%bfcl prompt |
| Relevance detection · BFCL% correct, higher is better 8 · 1 of 81 measured | 93.8%bfcl prompt | 100.0%bfcl prompt |
| Single-turn calls (curated) · BFCL% correct, higher is better 63 · 74 of 81 measured | 79.4%bfcl prompt | 61.1%bfcl prompt |
| Single-turn calls (user-contributed) · BFCL% correct, higher is better 37 · 64 of 81 measured | 74.2%bfcl prompt | 60.8%bfcl prompt |
| Web search · BFCL% correct, higher is better 46 · 56 of 81 measured | 4.0%bfcl prompt | 1.0%bfcl prompt |
Quick answers
Is Gemma 3 12B better than Gemma 3 4B?
They score about the same, so pick Gemma 3 4B: it costs 43% less per token.
Which is cheaper, Gemma 3 12B or Gemma 3 4B?
Gemma 3 4B is cheaper: $0.042 against $0.075 per million tokens, 43% less.
Which is better for agents and tool use, Gemma 3 12B or Gemma 3 4B?
Gemma 3 12B ranks 144 of 159 models on sale for multi-step tasks with tools: support desks, coding agents, function calls, against 158 for Gemma 3 4B.
Which is better for reasoning and knowledge, Gemma 3 12B or Gemma 3 4B?
Gemma 3 4B ranks 181 of 187 models on sale for hard questions across science, maths and general knowledge, against 187 for Gemma 3 12B.
Which is better for writing, Gemma 3 12B or Gemma 3 4B?
Gemma 3 12B ranks 137 of 169 models on sale for what people prefer in blind comparisons, and judged writing quality, against 157 for Gemma 3 4B.
Which is better for sticking to the facts, Gemma 3 12B or Gemma 3 4B?
Gemma 3 12B ranks 6 of 139 models on sale for summarising without inventing things, and factual answers, against 32 for Gemma 3 4B.
Run Gemma 3 12B and Gemma 3 4B on your own prompt.
Benchmarks aren't your task. Test both side by side on your prompt, with every answer checked against your rules and the cost of each one.