Gemma 4 E2B and E4B Explained

The “E” means effective parameters, not total. Gemma 4 E2B holds 5.1B parameters but behaves like a 2.3B model at inference; E4B holds 8B and behaves like 4.5B. The gap comes from Per-Layer Embeddings, and it changes how you budget memory.

What does the E in E2B and E4B mean?

It is a claim about compute, not about storage. Per-Layer Embeddings give each decoder layer its own embedding for every token, so a large share of the checkpoint is embedding tables that are looked up rather than multiplied through. The effective count is what remains in the active compute path.

Model	Total parameters	Effective parameters	Context	Catalog value
Gemma 4 E2B IT	5.1B	2.3B	128K	gemma-4-E2B-it
Gemma 4 E4B IT	8B	4.5B	128K	gemma-4-E4B-it

Both cards list text, image, and audio as supported input modalities, which is unusual at this size — the other students in the distil labs catalog are text-only.

How much memory do you actually need?

Two numbers, and you need both. Weight storage follows the total count; the fast-path working set follows the effective count. At BF16, two bytes per parameter:

Model	Total × 2 bytes	Effective × 2 bytes
E2B	5.1B × 2 = 10.2 GB	2.3B × 2 = 4.6 GB
E4B	8B × 2 = 16.0 GB	4.5B × 2 = 9.0 GB

These are derived from the parameter counts on the model cards, not measured on hardware. The practical reading: your disk and download budget is the left column, and neither model is a 4 GB-VRAM candidate at BF16 — see which SLM fits in 4GB of VRAM for the full arithmetic and what quantization does to it.

Why would you pick E4B over E2B?

Because you need the headroom and can pay for it. E4B roughly doubles both numbers above. Nothing in the distil labs published benchmarks separates them — our 12-model and 15-model studies predate the Gemma 4 release, and the mid-size MoE benchmark tested gemma-4-26b, a much larger sibling.

So the honest answer is: train E2B first, read the metrics, and step up only if you miss your bar. That ordering is cheaper than reasoning about it, and base-model accuracy is a poor predictor of fine-tuned accuracy anyway.

What are the downsides of the effective-parameter design?

The marketing number is the smaller one. “E2B” reads like a 2B model in a catalog listing, and it is not one on disk. If you size a deployment from the name, you will be short by more than 2x.

Multimodality you may not use. Image and audio encoders are parameters you ship and store whether or not your task has images or audio in it. For a text classifier, that is dead weight.

No published fine-tuning track record. Gemma 4 is new enough that there is no distil labs rank for it, and the older Gemma 3 line did not top any of our rankings — the 12-model study put the Qwen3 family in four of the top six fine-tuned spots and noted Gemma trailing.

128K context is not free. As with any long-context model, KV cache grows with sequence length, so the usable window on constrained hardware is far below the advertised one.

Can you use Gemma 4 for tool calling on distil labs?

Yes, and this is what most distinguishes it from Gemma 3. The supported models catalog limits tool-calling tasks to the Qwen3, Qwen3.5, Llama 3, LFM2/LFM2.5, FunctionGemma, and Gemma 4 families. Gemma 3 is excluded at every size; Gemma 4 is not.

base:
  task: tool-calling-closed-book
  student_model_name: gemma-4-E2B-it
  teacher_model_name: zai.glm-5

synthgen:
  teacher_temperature: 0.6

Remember that the teacher also has to support tool calling — the catalog marks that with a tick. openai.gpt-oss-120b is the documented default and qualifies; GLM-5 is the teacher recent distil labs work reaches for, also qualifies, and needs the 0.5-0.7 temperature band that applies to every reasoning teacher. Which teacher model should you pick covers the full choice. If you want the smallest thing that calls functions rather than the most capable, compare against FunctionGemma 270M and LFM2.5 350M before committing to a multi-gigabyte checkpoint.