Qwen vs Llama vs Gemma for Tool Calling
Qwen and Llama are eligible for tool-calling tasks across their whole small-model range. Gemma is not: only Gemma 4 and the purpose-built FunctionGemma qualify, and Gemma 3 is excluded at every size. Within the eligible set, fine-tuning moves accuracy far more than family choice does.
Which families can you even use for tool calling?
Eligibility is a hard gate, not a preference. The supported models catalog limits tool-calling-closed-book and multi-turn-tool-calling-closed-book to a named subset of students.
| Family | Eligible students | Excluded |
|---|---|---|
| Qwen | Qwen3 (0.6B-8B), Qwen3.5 (0.8B-9B) | — |
| Llama | Llama 3.2 1B/3B, Llama 3.1 8B | — |
| Gemma | Gemma 4 E2B/E4B, FunctionGemma 270M | all Gemma 3 sizes |
| Liquid | LFM2, LFM2.5 | — |
| SmolLM2 | none | 135M and 1.7B |
Picking an ineligible student is a configuration error, not a quality trade-off — training will not run. If you are on Gemma 3 today and tool calling is on the roadmap, that migration is a base-model change, so make it before you invest in data.
How do the three families differ once they are trained?
Less than you would expect, on the tasks where we have measurements. Every result below comes from fine-tuning on distil labs with the same pipeline and a GPT-OSS-120B teacher, scored on tool call equivalence.
| Model | Family | Task | Base | Tuned | Teacher |
|---|---|---|---|---|---|
| FunctionGemma 270M | Gemma | Shell (Gorilla) | 9.90% | 96.04% | 97.03% |
| FunctionGemma 270M | Gemma | Smart home | 38.82% | 96.71% | 92.11% |
| FunctionGemma 270M | Gemma | Voice assistant | 23.35% | 90.86% | 96.95% |
| LFM2.5 350M | Liquid | Shell (Gorilla) | 61.4% | 98.0% | 97.03% |
| LFM2.5 350M | Liquid | Smart home | 63.2% | 96.7% | 92.11% |
| LFM2.5 350M | Liquid | Voice assistant | 34.5% | 95.9% | 96.95% |
Two honest gaps in that table. We have not published per-task tool-calling numbers for Llama, and the Qwen figures we have are from the 15-model study, where fine-tuned Qwen3-4B beat the teacher by 0.04 on smart home and 0.03 on the voice assistant. Anyone presenting a clean three-way tool-calling leaderboard for these families is presenting numbers we did not measure.
What the table does establish: base scores vary enormously between families — 9.9% to 63.2% on the same tasks — and tuned scores converge into a 90-98% band. The starting point is a family property; the finishing point mostly is not.
Does model size matter more than family?
For eligibility, no. For results, less than fine-tuning does. A 270M FunctionGemma reached 96.04% on shell commands, within a point of a teacher 445 times its size.
Family still shows up in two places. The Qwen3-1.7B card states the model “excels in tool calling capabilities” and ships agent integration, so its base scores start higher — useful if you cannot fine-tune at all. And Llama’s advantage in our 12-model benchmark is tunability: Llama-3.2-1B-Instruct ranked first at 3.44 and Llama-3.2-3B second at 4.67, meaning the family gains most from training. That is the property you want when your seed data is thin.
Which teacher do you need for a tool-calling task?
One marked with a tick in the catalog’s tool-calling column — the constraint applies to teachers as well as students. GPT OSS 120B and 20B (both plain and thinking), DeepSeek V3.2, Qwen3 235B A22B, GLM 5, Kimi K2 Thinking, Kimi K2.5, and MiniMax M2 Thinking qualify. DeepSeek R1, DeepSeek V3.1, Qwen3 480B Coder, and Qwen2.5 VL 72B do not.
openai.gpt-oss-120b is the documented default, qualifies, and is the teacher behind the measured numbers in the table above. It is not the only sensible pick: recent distil labs work uses GLM-5 (zai.glm-5, 744B total with 40B active), and Kimi K2.5, Kimi K2 Thinking, MiniMax M2 Thinking, and DeepSeek V3.2 are all tool-calling-capable teachers too. Every one of those is a reasoning model, so synthgen.teacher_temperature has to sit between 0.5 and 0.7. Full guidance is in which teacher model should you pick.
So which family should you choose?
Choose Qwen if you want one base model that covers classification, QA, and tool calling, and you may not fine-tune every task. It has the best published all-round fine-tuned record and the strongest untuned tool-calling behaviour of the three.
Choose Llama if your seed data is thin and you are relying on training to close the gap. It is the most tunable family we have measured.
Choose Gemma only as FunctionGemma or Gemma 4. FunctionGemma is the right answer when the deployment target is a phone or an embedded device and the task is nothing but function calls — see FunctionGemma 270M for multi-turn tool calling. Gemma 4’s effective-parameter design is explained in Gemma 4 E2B and E4B explained.
Consider Liquid even though it is not in the title. On our three multi-turn benchmarks, LFM2.5 350M posted the highest tuned scores of any model we have published at that size.