2 min read

Qwen3.5 4B: almost Gemma, for 2.2GB less

Why a fourth Qwen entered the catalogue, where it sits on the Pareto frontier, and which model to pick for your VRAM.

Qwen3.5WebGPUAgent

We resisted adding a fourth Qwen for a long time. A catalogue of five models is already a lot to explain, and each new line is a choice that the user must make before having tried anything. Then we measured Qwen3.5 4B against Gemma 4 E4B, and the argument became difficult to refuse: almost the same quality, for 3GB instead of 5.2GB.

The numbers

BudgetValue
Download~3.0GB
Working set~4.4GB
Minimum GPU buffer1.5GB
Precisionq4f16
Runtimeqwen3_5

The 2.2GB saved on the download are not a detail. They are the difference between a model that one tries and a model about which one thinks "later". And the buffer at 1.5GB instead of 2GB opens the door to a whole family of GPUs which were excluded.

Where it is placed

Our catalogue is a Pareto frontier over (size, capability). A model enters only if nothing smaller does as well, and it leaves the day something smaller catches it up. Qwen3.5 4B entered without discussion:

  • above Qwen3.5 2B on everything that asks for reasoning over several steps, and clearly above on long code;
  • below Gemma 4 E4B, but by a margin which does not justify 2.2GB more for the majority of uses.

It shares the qwen3_5 runtime with the 2B, so it benefits from the same KV cache between turns. On an agent session which accumulates tool results, this is what makes the difference between a usable loop and a loop where one waits twenty seconds at each step.

Sampling

Unchanged, the family preset:

{ "temperature": 0.7, "top_p": 0.8, "top_k": 20 }

Which one to take, in the end

The question comes back often, so here is our answer without diplomacy:

  • 4GB of VRAM or less — Qwen3.5 2B.
  • 6 to 8GB — Qwen3.5 4B. This is the best ratio of the catalogue.
  • More than 8GB and you want the maximum — Gemma 4 E4B.
  • An old machine, or you just want to see — LFM2.5 350M, 255MB, it loads everywhere.

If you do not know which category you fall into, the AI Readiness Test answers in a few seconds.

Qwen3.5 4B is available in the Local LLM workspace and in the Agent.

Read next