NVIDIA retrained Gemma-4 again, and its 4B active parameters squeeze to the front of the benchmark chart

NVIDIA has released the BF16 version of Nemotron-3.5-Bolt-E4B on Hugging Face. It has 8B total parameters and 4B active parameters; the base model is Gemma-4-E4B-it, and NVIDIA continued post-training on top of it. The model card is very direct: it focuses on improving multi-turn tool calling, instruction following, and input grounding, targeting local voice assistants, device assistants, and embedded interfaces with low memory and low TTFT.

This chart is the core showcase in the model card. The horizontal axis is time to first token, the vertical axis is overall performance, and the color and size of the points correspond to memory usage. Nemotron-3.5-Bolt-E4B 4GB lands at 70.2, Gemma 4 26B-A4B is at 72.8, and Qwen 3.8 27B is at 67.0. The chart is called Smart and Fast, and the point is clear: it wants to prove that a small model can also carve out a position between response speed and score.

The accompanying tooling is also fairly complete. DFlash drafter is used to accelerate decoding, and ggml and unsloth quantized versions are also available. The Q4_K_M quantized version is called NVIDIA-Nemotron-3.5-Bolt-E4B-4GB, and the model card directly says it is at the 4GB VRAM level.

But you cannot judge benchmark results by a single 'outperforms' line. In the model card benchmark table, Nemotron-3.5-Bolt-E4B scores 53.10 on MultiChallenge, Gemma4 E4B-it scores 42.08, and Qwen3.5 4B scores 51.76. RoleBench is 67.05 versus 49.95 versus 53.66. IFBench is 63.79 versus 42.84 versus 58.91. It also has advantages on RoleBench, When2Call Multi-Turn, and IHEval, which are related to tool calling and multi-turn dialogue.

Qwen3.5 4B is no pushover. On Tau2-Bench, Qwen3.5 4B scores 75.06 and Nemotron scores 55.10. On RULER 64K, Qwen3.5 4B scores 95.15 and Nemotron scores 86.99. On LongBench v2 48K, Qwen3.5 4B scores 59.93 and Nemotron scores 50.50. So the more accurate statement is that NVIDIA pushed the Gemma-4 base a step further toward tool calling, multi-turn interaction, and local assistants, leading on some items while still being held down by Qwen3.5 on long context and certain agent benchmarks.

One commenter joked in the replies: '8B parameters, 4B active—just like me at the Friday stand-up.' Others asked how the 4B active parameters actually perform in long agent loops. The model card provides 64K input/output context and a bunch of benchmarks, but does not give empirical conclusions for long loops; that question can only be filled in by people who actually deploy it.

There is one deployment detail worth noting. DFlash speculative decoding follows the vLLM and NVIDIA GPU route. One commenter mentioned that DFlash cannot be used on Mac, but in optiq you can use the Gemma-4 E4B assistant drafter; the tokenizer is the same, and the speedup is noticeable when running code. This claim comes from a commenter's testing, not a promise from the model card. The model card also shows that this BF16 version is mainly intended for customization, post-training, quantization, and evaluation, not for taking directly into production inference. If you really want to run it locally, the Q4_K_M 4GB version is more practical.

My take: The truly interesting part of this release is that NVIDIA took a mature small base model, did targeted post-training, and then laid out the drafter, quantization, and deployment path together. The parameter scale is not intimidating, but it targets the real-world scenarios of local voice and device assistants. If you want an all-around long-context champion, Qwen3.5 still leads on multiple items. If you want a small model that can fit into a device, responds quickly, and has decent tool-calling scores, Nemotron-3.5-Bolt-E4B deserves a place on the shortlist.