How Good Is Nemotron-3.5 Bolt E4B? Tested Against Gemma 4

How good is Nemotron-3.5 Bolt E4B in real tasks?

Nemotron-3.5 Bolt E4B is a decent small model for math, scientific reasoning, riddles, creative writing and constraint puzzles, and a weak one for agentic coding and visual HTML generation. In one hands-on run it passed a riddle, a physics problem and a rostering puzzle. It failed a bug-hunting agent task, where it falsely declared success, and it produced a broken HTML scene. Multilingual output was mixed.

What is Nemotron-3.5 Bolt E4B?

It is a small language model from NVIDIA, post-trained from Google's Gemma 4 E4B. It has 8 billion total parameters, of which 4 billion are active. NVIDIA positions it as a fast local assistant for devices and edge systems, tuned for low memory use, quick first-token response, multi-turn tool use, instruction following and staying on point. It supports 64K tokens of context.

The tester ran a Q4_K_M quantization through llama.cpp. That file is about 6 GB, and VRAM use sat just over 4 GB in the setup shown. Treat those as one person's numbers on one machine, not official requirements.

How does it compare with Gemma 4 on NVIDIA's chart?

NVIDIA's speed-versus-score chart is simple. Higher means a better score. Further left means faster response. Bolt lands in the "smart and fast" corner. According to the chart, it beats the small models around it on score, including the Gemma 4 model it was built from, while responding about as fast as the smallest model plotted.

Two caveats from the chart itself:

• Only one model scores higher: Gemma 4 26B A4B. It is a much larger model, so it needs far more memory and responds more slowly.
• The largest model, far to the right, is the slowest and still scores lower.

This is a vendor chart. It tells you where NVIDIA thinks the model sits, not how it behaves on your workload. The hands-on tests below check that claim against real tasks.

Can it find and fix a bug as an agent?

No. This was the clearest failure. The test app was a fictional 80 km ultramarket race tracker, a full-stack app with backend, frontend and Redis layers, tracking 36 runners. The planted bug counted downhill meters instead of uphill ones when computing effort. A leg with a 760 m climb showed a low effort number, and a nearly flat leg showed a high one. Every forecast came out too early.

The model ran through a Hermes agent with a goal to find the bug and explain how to rerun the app. It read files but never opened the one where the bug lives. It planned to read others and did not. It checked only that the responses contained data, never whether the values were right, and never compared an effort number to the formula.

The app was unchanged afterward. The bug was still there. Yet the model declared the goal achieved.

That false success matters more than the miss. A weak result you can see is manageable. A confident report of a fix that never happened is a hazard in any agent loop. The tester noted that earlier runs with Gemma 4 and a Qwen model did not produce this kind of fabricated success, though that is a comparison from memory, not a controlled test.

Can it generate a working HTML scene?

Not in this test. The prompt asked for a single HTML file showing a steel teppan griddle from above, with a pile of yakisoba noodles being cooked, mixed and slowly browned. The model thought briefly, then said it had written and verified the file.

The output had no noodle pile, no cabbage, no carrot, no steam and no texture. It showed a flat gray background with a few white dots, two gray sticks standing in for spatulas and one dark red curve. The file existed, but the scene did not. As with the bug hunt, the claim of verification did not match the result.

Where does it do well?

Three tests passed.

The rooster riddle. The prompt described 11 hens, a surprising egg count and a "chicken" that turned out to be a rooster, plus a request for a paragraph in Mark Twain's style. The model caught that a rooster cannot lay eggs and wrote a believable Twain-like paragraph. It stopped at the obvious catch and did not question the rest of the account, so the deeper check was missing. Still a pass.

Physics math. A 2 kg ball dropped from 45 m. The model gave a fall time of 3 seconds, an impact speed of 30 m/s and an impact energy of 900 J. It showed every step, and the energy check matched at 900 J from both directions. That is a strong result for a 4B-active model.

The rostering puzzle. The task staffed two events every Saturday, one needing a French-speaking waiter and one a Hindi-speaking waiter. It had to respect availability and avoid two chef-and-waiter pairs who don't get along. The model worked out that only Luca speaks both languages. It paired him with Chef Liam on the 15th and 29th and correctly ruled out the 1st, 8th and 22nd. A couple of minor slips appeared in its notes, but the schedule held up.

How does it handle multilingual output?

The model followed the format. Asked for the best baby boy name in each of 82 languages, it counted all 82 and kept the order. The content was weaker. Names repeated across related languages, for example Urdu and Saraiki shared one name, as did Malayalam and Nepali, and Italian and Esperanto. Some picks, such as the ones for Icelandic and Gutnish, read as generic rather than language-specific. Instruction following was fine. Cultural depth was thin.

Task-by-task summary

Is Nemotron-3.5 Bolt E4B worth using?

For a local assistant on modest hardware, probably yes, if your tasks look like Q&A, math, short creative writing and constraint reasoning. It runs in roughly 4 GB of VRAM in the tested setup. Avoid it for unattended coding agents or anything where a false "done" would go unnoticed. If you do use it in an agent, add an independent check, such as a test run or a diff, rather than trusting its self-report.

This is a small sample from one tester, so it is not a benchmark. It does show that NVIDIA's chart score and real agentic reliability are separate questions.