NVIDIA's 4 GB Nemotron Bolt Beats Its Old Model by 40%. On the Public Multi-Turn Bench, Qwen 3.5 4B Beats It by 20 Points.
NVIDIA released Nemotron 3.5 Bolt E4B 4GB on October 23: a Gemma 4 E4B derivative with Nemotron post-training, a Q4 GGUF around 4 GB, a DFlash drafter, and an OpenMDW-1.1 license. NVIDIA claims about 40% higher average quality than Nemotron 3 Nano 4B and up to 80% fewer output tokens than peers. NVIDIA's own table shows Tau2-Bench at 55.1 against 75.1 for Qwen 3.5 4B, and the benchmarks where Bolt wins, Brevity Bench and When2Call-MT, are NVIDIA's own and not yet public. We read NVIDIA's research page and the Hugging Face card.
By T.W. Ghost
NVIDIA released Nemotron 3.5 Bolt E4B 4GB today, October 23, 2026: a small conversational and agent model built on Google's Gemma 4 E4B with Nemotron post-training, shipped as a Q4_K_M GGUF of about 4 GB with a DFlash speculative drafter, under the OpenMDW-1.1 license. NVIDIA's claims: about 40% higher average quality than its own Nemotron 3 Nano 4B, and up to about 80% fewer output tokens than peer models.
The comparison that matters is in NVIDIA's own table. On Tau2-Bench, the public multi-turn customer-service benchmark, Bolt scores 55.1 and Alibaba's Qwen 3.5 4B scores 75.1. The benchmarks where Bolt wins, Brevity Bench and When2Call-MT, are NVIDIA's own, and NVIDIA describes them as to be released. We read NVIDIA's research page and the Hugging Face card. Here is what held up.
What You Are Downloading
| Nemotron 3.5 Bolt E4B 4GB | |
|---|---|
| Released | October 23, 2026 |
| Base | Gemma 4 E4B (Gemma4-E4B-it) with Nemotron post-training |
| Format | Q4_K_M GGUF, about 4 GB; DFlash drafter |
| License | OpenMDW-1.1 (per the Hugging Face card); Gemma terms apply to the base |
| Parameter count | Not stated as a plain number; "E4B" is Gemma's effective-4B naming; an "8B" figure on X is unverified |
| Vendor claims | About 40% above Nemotron 3 Nano 4B average quality; up to about 80% fewer output tokens vs peers |
| AA entry | None found |
"4 GB" is the quantization, not the architecture. The size describes the Q4 GGUF at a typical batch-one footprint. The model underneath is a Gemma 4 E4B derivative.
Read two licenses. OpenMDW-1.1 governs NVIDIA's release; the base is Gemma. Check both before you redistribute.
The Benchmarks, All of Them NVIDIA's
| Benchmark (NVIDIA's table) | Bolt E4B | Nemotron 3 Nano 4B | Qwen 3.5 4B | Gemma 4 E4B |
|---|---|---|---|---|
| Tau2-Bench (multi-turn customer) | 55.1 | 29.2 | 75.1 | 38.5 |
| Brevity Bench (NVIDIA's, to be released) | 82.5 | n/a | 75.3 | 80.0 |
| When2Call-MT (NVIDIA's, to be released) | 63.4 | n/a | 37.8 | n/a |
Three honest observations:
- Qwen wins the public test by 20 points, on NVIDIA's own table. Tau2-Bench is the one benchmark here anyone can run, and it is the one Bolt loses.
- The wins are on tests nobody else can run yet. Brevity Bench and When2Call-MT are NVIDIA benchmarks described as not yet released. A win on your own private test is a claim, not a result.
- The 40% is against NVIDIA's previous small model. Beating Nemotron 3 Nano 4B by 40% is consistent with still trailing Qwen 3.5 4B by 20 points, because Nano scored 29.2 on Tau2.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 23; Gemma 4 E4B base; Q4 GGUF about 4 GB; DFlash drafter; OpenMDW-1.1 | Verified (NVIDIA research page; Hugging Face card) |
| About 40% over Nemotron 3 Nano 4B; up to 80% fewer output tokens | Vendor |
| Tau2-Bench 55.1 vs Qwen 3.5 4B 75.1 | Vendor table (NVIDIA's own numbers) |
| Brevity Bench and When2Call-MT wins | Vendor, on NVIDIA benchmarks not yet public |
| Parameter count "8B" | Unverified (X only) |
| Artificial Analysis score | Unverified (none found) |
Who Should Use It
Try Bolt if you want a 4 GB conversational model with a drafter and you can benchmark it on your own multi-turn tool tasks; its design goal, fewer output tokens, is worth measuring on your traffic.
Skip the marketing comparison: on the public multi-turn test NVIDIA chose to print, Qwen 3.5 4B is the better model by a wide margin, and the tests where Bolt wins cannot be checked yet. Read OpenMDW-1.1 and the Gemma terms before shipping.
The frontier is elsewhere: at 4 GB nobody expects a frontier score, and Artificial Analysis has not scored it at all. Our model comparison has the current standings side by side.
Sources
- NVIDIA Research: Nemotron 3.5 Bolt E4B 4GB
- Hugging Face: nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB
- OpenMDW license 1.1
*Not sure whether a 4 GB model is enough for your agent's conversations? Take the free 2-minute quiz and get matched. Then read about the laptop-class coding model released two days earlier under a license with fewer questions: Recursive Flint.*