A 4GB model built on someone else’s
Nvidia has released Nemotron 3.5 Bolt E4B 4GB, a small open model aimed at voice assistants, game characters and robots that run on the device rather than in a data centre. The model’s own page says it “builds on Gemma 4 E4B with Nemotron post-training” — the base is Google’s Gemma4-E4B-it, and what Nvidia contributed is the post-training on top of it.
That is worth stating plainly because the release is being read as an Nvidia model. It is Nvidia’s post-training recipe applied to a Google base, which is what the open-weights ecosystem now looks like from the inside: the companies that sell the hardware increasingly ship models derived from the companies that do not.
The context window is 64k tokens. The model supports 35 languages, on a base Nvidia says was pre-trained on “+140 languages.”
What the 4GB in the name means
Not system memory. The figure refers to the quantised weights: Nvidia’s benchmarks are run at BF16, while the GGUF release is Q4_K_M, and the page says the model fits “as little as ~4GB in its target workloads.”
Nvidia’s own table is more specific than the headline number. On an RTX 5070 with 12GB, it reports 3.5GB at a short context and 4.3GB at 32k context with 2,000 output tokens. On a Jetson Orin Nano with 8GB, the same two configurations take 4.2GB and 4.9GB. Throughput at batch size one is 133.4 tokens a second on the RTX 5070 and 15.5 tokens a second on the Jetson Orin Nano — a nine-fold difference between the two devices Nvidia is pitching this at.
The benchmark table does not say what the headline says
Nvidia publishes 21 benchmark rows comparing Bolt E4B against its own Nemotron 3 Nano 4B, Alibaba’s Qwen 3.5 4B and the Gemma 4 E4B it was built from. These are Nvidia’s own numbers, not an independent evaluation.
Bolt leads most of them, sometimes by a lot. On RoleBench it scores 67.0 against Nano 4B’s 12.0. On IHEval it scores 84.8 against 51.3. On When2Call-MT it scores 63.4 against 27.7. Against the Gemma 4 base it started from, it gains more than 20 points on IFBench and nearly 18 on IHEval.
But the same table shows Qwen 3.5 4B ahead on six rows, and two of the gaps are not small. On Tau2-Bench, a multi-turn tool-use test, Qwen scores 75.1 and Bolt scores 55.1 — a 20-point deficit. On LongBench v2 at 48k, Qwen scores 59.9 to Bolt’s 50.5.
The row worth dwelling on is RULER at 64k, the long-context recall test run at exactly the context length Nvidia advertises. Bolt scores 87.0. Nvidia’s own Nemotron 3 Nano 4B scores 91.5, and Qwen 3.5 4B scores 95.1. Bolt beats only the Gemma 4 base it was post-trained from, at 86.5, and places third of the four models in its own comparison at its own headline context length.
Nvidia’s summary claims are about averages rather than individual rows: roughly 40% higher average quality than Nemotron 3 Nano 4B, roughly 10% higher than the other models evaluated, and “up to ~80% fewer tokens” in comparisons where reasoning is enabled. Both of those can be true while the long-context row goes the other way.
Where to get it
Weights are on Hugging Face as a GGUF post-trained model and a BF16 checkpoint, with DFlash drafters for the GGUF and NVFP4 builds. The page also lists Ollama, LM Studio, Unsloth, Canonical, Jetson AI Lab and Nvidia’s In-Game Inferencing SDK as distribution routes. The page does not name a licence.
What to watch
The useful test is whether anyone reproduces the table. Bolt’s strongest claims are on conversation, persona and instruction-following benchmarks, which are the ones most sensitive to harness and prompt choices, and the weakest result is on RULER, which is the most mechanical of them. If independent runs confirm the persona gains and the long-context gap, Nvidia has shipped a model that talks well and recalls poorly — a reasonable trade for a game character, and a bad one for an agent.
