Gemma 3 270M: What It Can Do
Gemma 3 270M is the smallest Gemma and one of the smallest students in the distil labs catalog. Fine-tuned, it is a credible classifier and extractor; untuned it is close to useless, and there are task types it cannot reach no matter how much data you give it.
How small is 270M in practice?
Small enough that the weights are not the constraint any more — the capability ceiling is.
| Property | Value | Source |
|---|---|---|
| Parameters | 270M | Gemma 3 270M IT card |
| Training tokens | 6 trillion | model card |
| Context length | 32,768 tokens | model card |
| Weights at BF16 (2 bytes/param) | 0.54 GB | 270M × 2 |
| Weights at 4-bit (~0.5 bytes/param) | ~0.14 GB | 270M × 0.5 |
| Catalog value | gemma-3-270m-it |
supported models |
Six trillion training tokens against 270M parameters is an unusually high token-to-parameter ratio — this is a heavily-trained tiny model, not a truncated large one. That is why it responds well to fine-tuning despite its size.
What can Gemma 3 270M do well?
Bounded classification and structured extraction, once trained. Tasks with a fixed, enumerable output space are where a 270M model spends its capacity efficiently: it has to learn a decision boundary, not a world model.
Our 12-model benchmark ranked it fifth on tunability at an average rank of 5.00, ahead of several models many times its size on gains-from-training. That ranking is about improvement, not final accuracy, and it is the honest reason to consider this model: it moves a long way when you train it.
The deployment story is the other half. At roughly 0.14 GB quantized it fits in places a 1B model does not — browser tabs, microcontrollers, a single CPU core alongside your application. Fine-tuning without a GPU is a realistic path at this size.
What can it not do?
Open-ended generation. The Gemma 3 card is explicit that the models “may generate incorrect or outdated factual statements” and can miss “subtle nuances, sarcasm, or figurative language.” At 270M those limits are not edge cases, they are the normal behaviour.
Tool calling on distil labs. The catalog restricts tool-calling tasks to the Qwen3, Qwen3.5, Llama 3, LFM2/LFM2.5, FunctionGemma, and Gemma 4 families. Gemma 3 is not on that list at any size. If you want function calling in this parameter range, the model to use is FunctionGemma 270M, which Google built for exactly that.
Closed-book QA over a large corpus. Facts have to be stored in weights, and there are not many weights.
Predictable results. Its tunability confidence interval in the 12-model study was ±2.77, the widest in that table. It gained a lot on some tasks and little on others. Plan for two runs, not one.
How does it compare with the other tiny students?
Three models sit within a hair of each other on size and diverge sharply on what they are for.
| Model | Parameters | Context | Built for |
|---|---|---|---|
| SmolLM2 135M | 135M | 8K | General text, English only |
| Gemma 3 270M | 270M | 32K | General text, task fine-tuning |
| FunctionGemma 270M | 270M | 32K | Function calling only |
Against SmolLM2 135M, Gemma 3 270M doubles the parameters and quadruples the context window; the 8K window is a real constraint once documents get involved. Against FunctionGemma, the choice is decided by task type rather than quality — same parameter count, different job.
Is 270M worth it over 1B?
Only when the size difference buys you a deployment you could not otherwise have. The 12-model benchmark’s practical recommendations put nothing from the Gemma family in the top tier for fine-tuned accuracy; Qwen and Llama take those spots. You are trading accuracy for footprint, deliberately.
The decision rule that survives contact with production: if a 1B model fits your hardware, run teacher evaluation and train the 1B first, because it is the higher accuracy ceiling. Drop to 270M when the 1B misses a latency, memory, or unit-cost target you actually have. Sizing logic in full is in what size model do you need.